 Vector databases have recently gained prominence with the rise of large language models and generative AI. A vector database is a data store for unstructured text in the form of vector embeddings for various AI models and applications. Embeddings are a high dimensional vector representation of text that conveys rich semantic information and represent an efficient way of capturing unstructured data like text.
The rising popularity of large language models like GPT-4, Gemini, Claude-2, Llama-2, Mixtral and others have fuelled tremendous interest in generative AI across the industry to build applications based on these models. Vector databases are specialized for handling vector data that is used to train or fine-tune these foundational models for domain and company specific use cases. Unlike traditional scalar-based databases, vector databases offer optimized storage and querying capabilities for vector embeddings. Although several vector databases are now available in the market like Pinecone, Chroma, Qdrant amongst others, deciding which vector database to choose for enterprise use cases is not a straightforward decision. In this article, you will learn how to decide which vector database to choose for your organization based on criteria like performance, reliability, scalability, cost-efficiency, developer experience, security, technical support amongst others. Key Considerations In this section, you will learn in detail about each of the key factors that should be considered to make your final selection of a vector database. These include data and use case characteristics, performance, functionality, enterprise-readiness, developer experience, and future roadmap. 1. Data and Use Case It is important to work backwards from the specific business use case that you are planning to solve by leveraging organizational data and the latest techniques from the field of generative AI. For instance, if your business objective is to build an enterprise knowledge management chatbot like McKinsey’s Lilli, you will need to organize and prepare all the in-house text data such as documents, emails, chat messages etc. The use case defines several aspects of the data, including its size, frequency, data type, growth in the volume of data over time, data freshness and consequently the nature of the underlying vector embeddings to be stored in the vector database. These vectors may be sparse, dense, and also span multiple modalities depending on the use case. Additionally, careful planning and scoping of the use case also helps you understand other crucial aspects such as the number of users, the number of queries per day, the peak number of queries at any given instant, as well as the query patterns of the users. Vector databases utilize indexing and vector search powered by k-nearest neighbors (kNN) or approximate nearest neighbor (ANN) algorithms. This empowers a vector db to perform similarity search and identify the most similar vectors in the database. This capability underlies enterprise use cases based on natural language processing such as question-answering, document analysis, recommender systems, image and voice recognition etc. 2. Performance 2.1 Query latency and query per second (QPS) The primary performance metrics of a vector db are the query latency, i.e., the time it takes to run a query and get the result and the query per second that defines the throughput in terms of the number of queries processed in a second. These parameters are critical for ensuring a seamless user experience for several applications that require real-time results such as chatbots. Typical QPS values range from ~50-300 and the average query latency from 25-100 ms depending on the underlying hardware. 2.2 Scalability Scalability measures the ability of the vector database to grow and expand further to support the requirements of its customers. The scale can be measured in terms of the number of embeddings that can be supported and in terms of horizontal scaling of existing resources and vertical scaling of additional servers. Typically, most existing vector db companies provide scale-out capabilities up to a billion vectors without any performance degradation. If the resources can scale automatically, then you can be rest assured that your application will always be up and running. 2.3 Accuracy A vector database is as good as its accuracy of retrieving the right set of results based on the user queries. Here, the choice of vector search algorithms to identify data sources with similar embeddings as the embedding of the user query is pivotal. There are several different algorithms used for powering vector search such as kNN, ANN, FAISS, NGT. These algorithms generate approximate results and the best vector databases provide a good trade-off between speed and accuracy. 3. Functionality 3.1 Filtering on metadata In practice, filtering vector search results based on the metadata helps reduce the search space, thus providing for faster and more accurate search results. Typical metadata includes information like dates, versions, tags and the ability of a vector database to store multiple metadata fields allows for a better search experience. 3.2 Integrations Integrating a vector database into the existing data and engineering infrastructure in your organization is critical to faster adoption and lesser time to value. The ability of vector databases to seamlessly integrate with essential infrastructure elements like the cloud infrastructure, underlying large language models, databases etc. is a key factor to consider. 3.3 Cost-efficiency While performance metrics and functionality are core to a technology, the cost should be reasonable and fit your budget. The pricing of vector databases is a function of the number of ‘write’ operations such as update and delete and the number of queries. Other factors that affect the cost include the dimensionality of the embedding, the number of vectors stored in the database, and the size of the metadata. Depending on your use case and requirements, it is essential to estimate the overall cost of running your application at scale on a monthly or quarterly basis and evaluate the overall costs relative to your budget and the expected revenue from running the AI applications. 4. Enterprise-readiness 4.1 Security and compliance For most enterprise companies, it is imperative that any external vendor they employ meets strict security and compliance requirements. These requirements include SOC2, GDPR, HIPAA, ISO compliance and others, depending on the domain in which the company operates. The data privacy and security standards have gone up in the light of recent cybersecurity attacks and breaches of customer data, and you should ensure that any vector db vendor meets your specific security and compliance requirements. 4.2 Cloud setup Several modern companies have undergone digital transformation and house their entire data and infrastructure in the cloud vs on-premise. You may choose to manage and maintain your infrastructure via a self-hosted setup or go for a fully managed SaaS platform. The benefit of a fully managed system is that it automates clusters with minimal requirements for you to provision and scale clusters or take care of operational issues. 4.3 Availability Availability, i.e. the ability of your vector db to run without any interruptions, issues or downtime is essential to not adversely impact user experience. Most vector database providers vouch for specific SLAs which should meet the requirements for your applications. Typical values include 99.9% for uptime SLA and a few hours to a few business days for response time SLA depending on the severity of the production issue. 4.4 Technical support More often than not, you might be stuck facing some issues with your vector db and need some hands-on support from the vendor to help troubleshoot the issue. Does the company provide you with a dedicated team who can be available at a short notice to get on a call and figure out how to solve the problem? The quality of responsiveness and customer support experience provided by a vector db company is valuable and helps you develop a stronger sense of trust in the company. 4.5 Open source vs Closed source Some vector db companies are closed source and operate under a proprietary license such as Pinecone. At the same time, there are a host of vector db companies that are open source under the Apache 2.0 license such as Qdrant or Chroma while also offering a fully managed service. This can also influence your choice of the vector db provider. 5. Developer experience 5.1 Community Software and AI engineers are the core professionals who will work on the vector db and integrate it in the company’s infrastructure and deploy your generative AI application to production. Therefore, the quality of experience that developers have with a vector db solution is integral in shaping your final decision. Having an open-source community on Slack or Discord helps build more engagement and trust with developers than commercial vendor support. It provides your developers an opportunity to learn from developers at other companies as well and discuss and solve issues by leveraging the wisdom of the community. 5.2 Onboarding Onboarding a new technology is challenging as it determines the time your developer team takes to properly understand the product, integrate it, troubleshoot any issues, and become an expert in using the vector database. The availability of APIs and SDKs as well as clear product demos and documentation goes a long way in reducing the barriers to understanding a new vector database so that your developers can build with speed and confidence. 5.3 Time to value Similar to the time to onboard a new vector db, another important factor is the time to business value. If a vector db provider vouches for a fast deployment of a production-ready application, then you can realize value sooner, and meet your business goals faster as well. A long gestation time from onboarding to business value is a deterrent for many fast-moving companies and startups especially in the current frantic race to adopt and ship generative AI applications. 5.4 Documentation The quality of the vector database’s documentation determines the time to onboard, time to value, and trust in the provider’s expertise and product. Clear instructions with tutorials, examples and case studies help your developers understand and master the vector db faster. 5.5 User education Similar to community-based offerings, expert technical content such as blogs, demos and videos focused on the existing as well as new features are helpful for your team to understand and build faster. In addition to text and video content, other offerings like user testimonials, workshops, conferences also help educate your team and build more trust in the vector db provider. 6. Future roadmap A final factor to consider is the product roadmap of the vector database provider. Vector databases are an emerging technology that will need to continuously evolve alongside the advances in generative AI models, chip design and hardware, and novel enterprise use cases across domains. Therefore, the vector db vendor should show the potential for evaluating long-term and future industry trends such as sophisticated vectorization techniques for a wider variety of data types, hybrid databases, optimized hardware accelerators for AI applications such as GPUs and TPUs, distributed vector dbs, real-time and streaming data based applications, as well as industry-specific solutions that might require advance data privacy and security. Conclusion Vector databases are an essential ingredient for modern generative AI applications built on unstructured data such as text. Their popularity has increased in parallel to the developments in the generative AI field such as large language models, large image models etc. to serve as the underlying database for handling high-dimensional data stored as vector embeddings. In this article, you learned about several important pillars to help your decision making about the choice of the vector database. These factors include data and use case considerations, performance-based requirements such as query speed and scalability, functionality requirements such integrations and cost-efficiency, enterprise-readiness including security and compliance, and developer experience including community and documentation. Several vector database companies have emerged to build this foundational infrastructure. There is no single ‘best’ vendor of vector db and the ultimate choice is highly contingent on your organization’s business goals. Therefore, a data-driven approach guided by the factors listed in this article will help you select the most optimal vector db for your organization.
Comments
Published by Ikigai Labs Introduction
Many types of business data are organized in time—for instance, customer purchases on an e-commerce website or frequent orders of inventory materials by companies. Making sense of this time series data is vital for data or business analytics teams to understand the future dynamics of consumption and demand for their companies' products and services. Therefore, building predictive models to forecast demand is a vital task. There's a whole range of statistical as well as machine learning (ML) models that can be leveraged to build business-critical time series forecasting applications. However, time series data can be highly variable, and no one time series forecasting model will be applicable across use cases. With recent progress in ML and deep learning, new models are being developed all the time that provide state-of-the-art forecasting performance. For instance, Amazon has been working on a series of time series forecasting models over the last decade to predict customer demand for its products, ranging from statistical models to random forests to deep learning models, and transformers. Similarly, your business can benefit immensely from leveraging time series forecasting models to make accurate predictions of customer demand. In this article, you'll learn about ARIMA, Prophet, and mSSa, three popular time series forecasting models. These models have proved to be highly robust, reliable, easy to understand and implement, and versatile for forecasting applications in industries such as e-commerce, finance, retail, and travel. By the end of this article, you'll have a better sense of which of these models might be best for your own use case. Why Do Time Series Forecasting Models Matter? Real-world time series data have several characteristic patterns that reflect the nature of consumption and demand. For instance, if you're in the business of selling electronic gadgets, it's important for you to know how much inventory to stock so that you can meet the number of customer orders. Demand for your products can also change over time due to factors such as seasonal variations, holidays, the weather, or special events like the launch of a new product. Therefore, accurately forecasting the dynamics of demand becomes a critical function for your business. Poor demand forecasts may lead to grave consequences such as a significant reduction in sales and revenue as well as losing market share to your competitors. Using time series forecasting models enables your company to predict demand for the next day, week, month, or quarter and helps you to plan and prioritize business objectives and strategy accordingly. The time series forecasting models that have emerged over the years are based on different assumptions about the nature of the underlying time series data; as such, they've been developed to suit specific applications. To determine the time series forecasting model that's right for you, you should start by conducting preliminary analytics and evaluating the statistical distribution and properties of your data. This is an important step in identifying the right set of algorithms to model your specific time series data. Getting the choice right can help make your process more efficient without the need to test out multiple models. Once you've set a good baseline in terms of your model's performance, you can further improve it by experimenting with its various parameters. Additionally, the right model allows you to place more confidence in the accuracy of its results. Therefore, defining the most relevant time series forecasting model for your specific business use case is an important decision. Choosing between ARIMA, Prophet, and mSSa As mentioned, your particular use case is a key consideration. You may have large amounts of historical data that can be leveraged to make demand predictions for the next day, week, or month. Predicting electricity demand is one example that fits this scenario. Maybe you don't have a lot of historical data but still need to make forecasts for functions like sales or viewership or usage of a particular feature or product. In this section, you'll learn about the underlying principles of the ARIMA, Prophet, and mSSa time series forecasting models and be able to decide which models would be better suited to your forecasting goals. The ARIMA Model Autoregressive integrated moving average, or ARIMA, is a forecasting algorithm based on the assumption that past time series data can be used to predict future values. The amount of past information to use for modeling is controlled by a hyperparameter, p. ARIMA also assumes that past forecast errors can also be used to improve forecasts. The most recent errors are indexed by another hyperparameter, q. ARIMA models are great for forecasting stationary time series data. This implies that the data does not contain any seasonal or temporary trends and the statistical properties of the source of the time series data, like the mean and variance, do not change over time. A time series can be made stationary through several methods, with the common technique being differencing, where each differencing value is the difference between the value at the current time period and the previous time period. The number of differences required to achieve stationarity is determined by a hyperparameter, d. ARIMA is widely used for demand forecasting use cases, such as predicting demand in food manufacturing, energy, or user demand for services like ride-hailing. The Prophet Model Prophet is an open-source time series forecasting package developed by the data science team at Facebook. It's available in both Python and R and has been widely adopted across key industries such as e-commerce, tech, and finance. The forecasting algorithm is based on an additive model that can be decomposed into three distinct components: trends, seasonality, and holidays. As the forecasting model can be decomposed into its constituent factors, it's easy to extract the model coefficients to understand the relative impact of seasonality, trends, and holidays on the forecast. Prophet is best suited for forecasting applications that are associated with:
Prophet is designed to make forecasting automated and efficient for business analysts who may not have specialized data science skills. Its default parameters often yield forecasts that are as accurate as those produced by experienced forecasters. It's easy to use by nonexperts and requires less hyperparameter tuning. The mSSa model Multivariate singular spectrum analysis, or mSS, is a novel time series forecasting method that was recently formulated by scientists at MIT; they've shown that on benchmark data sets focused on time series data from electricity grids, traffic patterns, and financial markets, mSSa performs competitively with state-of-the-art neural networks for time series, such Amazon's DeepAR and LSTM. mSSa is particularly useful for modeling multiple time series with a varying number of observations per time series; it's also highly effective at imputation, or filling in missing values. mSSa has also been used to predict real-time traffic flow in software-defined networks with high levels of accuracy. Conclusion Forecasting demand is key for businesses to respond to fluctuating customer demand for their products and services. In this article, you learned about three popular time series forecasting models that are based on different statistical foundations: ARIMA, Prophet, and mSSa. These models have been used extensively at both startup and enterprise organizations, and you're now better equipped to choose which one could be right for you. Time series forecasting models can be built from scratch using libraries in R, Python, etc. Alternatively, for some organizations, it makes more sense to leverage existing platform solutions. For example, Ikigai provides a forecasting solution that includes all available algorithms including ARIMA, Prophet, mSSa, linear regression, etc., that can be easily configured using its no-code interface. When analysts are not sure which model to use, they can easily compare different ones with a one-click interface, or rely on AutoML to help them select the best model for their specific data. Additionally, Ikigai also provides a proprietary forecasting method called DeepCast that uniquely leverages statistical models with additional layers of machine learning on top of it, resulting in 20% more accurate forecasts vis-a-vis other state-of-the-art methods. Further, DeepCast is capable of making an accurate prediction based on only three weeks of data. Introduction
Data is the cornerstone of businesses from large enterprises to small D2C brands, and huge amounts of it can be collected from websites, mobile apps, chat messages, call centers, business transactions, surveys, and social media platforms, among other channels. All this data represents a gold mine of information that can offer customer insights and lead to new ideas for features or products. However, making sense of the data is easier said than done. The information originates from various channels and in multiple formats. It can be logged erroneously and contain other errors, including missing values. Because it comes from multiple domains, it can include unstructured data like text, images, audio, and video. That is why data preparation is essential. This involves cleaning, curating, transforming, and storing data sets for downstream applications including data analytics and data visualization, as well as predictive intelligence based on machine learning and deep learning models. Data can only provide value once it has been processed from its raw form, and effective data preparation can maximize that value. This article will explain the process of data preparation, especially in terms of data labeling, and will provide a checklist for data engineers to follow. What Is Data Preparation? Data preparation is not an entirely new process in technology companies. Data-driven operations previously focused on statistical analysis of business data from structured tables. The deep learning model has grown over the past decade along with the global penetration of mobile phones, widely available internet access, and cheaper cloud storage facilities. Today an estimated 2.5 quintillion bytes of data are being generated daily. Every user interaction with online companies is recorded, from someone clicking an ad or adding a product to a shopping cart to sharing a photo on a social media app. User-generated data is generally unstructured data: images, text, audio, or video. Such data can be used to train sophisticated deep learning models to predict what users want to type in a text, which branded products are featured in an image, and what kind of customer service will be provided in a phone conversation. For deep learning models to make sense of this data, all data samples need to be labeled. Data labeling tells the machine learning models what knowledge they need to acquire via supervised learning to power smart applications. This makes labeling critical in preparing data sets for training machine learning models. However, data labeling can also represent the chief source of errors, affecting potential improvement in model performance. Machine learning models can only be as accurate as the labeled data, which represents the models’ entire knowledge for the particular use case. For example, the source image data set in a face recognition program requires a label for every face shown in every image. During the labeling process for this data set, every image is reviewed by human subject matter experts, crowdsourced labelers on platforms like Amazon Mechanical Turk, or algorithms. Labeling helps clean and prepare the data set by removing noisy or unusable data. In this case, images that don’t contain any faces, or that show unreadable faces due to poor lighting or angles, should be removed because they won’t be helpful in training a face recognition model. This step also ensures the inclusion of images that are most helpful for the desired use case. Once the data set is reviewed and annotated, it can be used for all subsequent face recognition applications instead of going back to the raw data set. This saves time and effort for data engineers, as well as data scientists who might build novel models using the same data set. Additionally, multiple labels and metadata can be applied to each image during the labeling process so that they’re available for new use cases. A tag that identifies the face as that of a man, woman, or child can be used for different computer vision applications. This can potentially give the data set more flexibility for the future. The labeling can be built upon in subsequent versions of the data set. Once the face recognition model is live in production, new images can be labeled to help the model overcome data drift and augment its performance in the face of changing data distributions. This continued labeling and organizing keeps the models more robust and consistent. Data Preparation Steps There are certain best practices to follow when preparing data sets for deep learning applications. Following is a checklist for data engineers when working with unstructured data: (1) Check data formats Samples in a data set, especially if collected via web scraping or crowdsourcing, may come in multiple data formats. For example, an image could be a JPEG, PNG, or TIFF, while an audio file could be a WAV, MP3, or FLAC. Check whether the data set samples are in different formats, so that you can standardize the format across all samples. (2) Verify data types Certain deep learning applications are based on multimodal data including text, images, audio, video, and structured metadata. For example, a model that predicts what video a user might watch next is trained using multiple data types. It verifies the type of each data sample, then indexes and stores them separately. Note that an individual data type like numbers might also belong to different types like int, float, or string. (3) Verify data dimensions It’s crucial to check the dimensionality of the samples in a data set. For example, a set of images containing faces may be gathered from different cameras, each associated with different default image dimensions. (4) Identify what data needs to be labeled Once you’ve completed the above steps, you can begin data labeling. It may not be feasible in some situations to label each data sample, because manual labeling can be prohibitively expensive and time-consuming. In this case, choose an appropriate number of data samples for labeling. For common machine learning classification use cases, you need to sample data for labeling from each category. (5) Determine what type of labeling to perform The same data sample can be labeled in multiple ways depending on the use case. For instance, an image containing people and cars may be labeled for faces, for segmenting people or cars, or for the vehicle registration plates. (6) Decide who will label the data Data labeling can be performed manually by domain experts, crowdsourced from non-experts, or done programmatically using rule-based or model-based algorithms. Determine which annotators will define what kind of data, depending on their expertise or level of training. If a data set will be labeled using software, then the required configuration parameters, protocols, and performance metrics need to be established so that labeling is consistent. (7) Review data for errors and mistakes Usually, the first round of data labeling contains errors. To improve the data quality and eradicate errors, more experienced annotators should conduct a second or third level of review. Depending on cost, time, and available resources, each data sample can also be independently labeled by multiple annotators; the most commonly provided label can be assigned as the final label. (8) Split the data set into training and testing segments Once a data set is labeled, split it into separate train and test subsets for training and evaluating the model, respectively. Depending on the use case and the amount of available data, the ratio might be 80:20, 90:10, or even 99:1. To obtain more reliable results, k-fold cross-validation is recommended. Multiple training and test sets are sampled randomly, and the final results are averaged across all the different folds. Conclusion Without the protection of systematic data preparation and labeling checks, you may find that poor quality data damages the accuracy and performance of any analysis or models based on that data. If you follow the above guide, you will be able to ensure your data is good quality and labeled accurately. Related Blogs
Published by Andela Introduction
Data culture refers to an organizational culture of using data to derive insights and make informed business decisions. Companies can build a strong data culture by arming themselves with data and the right set of people, policies, and technologies. A data culture helps companies become more competitive and resourceful by leveraging data. And data-driven companies make better, faster, and more objective business decisions. They promote greater employee engagement and retention, and drive better financial outcomes in terms of revenue, profitability, and operational efficiency. In this article, you'll learn about data culture, what its importance is for modern organizations, and how you can build a strong data culture at your company. Why You Need a Strong Data Culture? Without a solid data culture, organizations will inevitably fail to harness the power of data. As previously stated, data culture refers to a set of beliefs and practices that companies use to cultivate and drive more data-driven decisions. Traditionally, businesses relied on the instinct and gut of a select few leaders to make strategic business decisions. However, with the accumulation and collection of massive volumes of customer and business data, domain expertise and instinct can now be complemented with data-driven insights to make more informed decisions. There are several advantages to building a strong data culture. Some of these include the following:
Every business sector, from product to finance to HR, creates and collects a lot of data from external customers or internal operations. For business heads and decision-makers, it's no longer feasible to stay on top of the ever-increasing volumes of data to better understand and evaluate the current state of their organization. However, with data analysts and scientists embedded across each department, it is possible to tap business insights in real time and respond quickly to changes in business performance. A strong data culture also promotes greater employee engagement and retention. When employees see that decisions are made on the basis of data and not driven just by the highest-paid executives, they feel that they can contribute more insights to influence decision-making. In the long term, this facilitates attracting the best talent in the market who can be incentivized to have a greater say in making key business decisions using data. Moreover, there are also strong financial outcomes associated with building and promoting a data culture. Companies with data-driven cultures benefit from increased revenue, better customer services, and more operational efficiencies leading to improved profitability. How to Build a Strong Data Culture? Building a strong data culture is a long-term endeavor that requires patient support and encouragement from leadership. Companies with strong data-driven cultures have executives who lead by example and establish clear expectations that decisions will be objective and based on data. Data leaders can lead from the front by establishing clear goals and guidelines, investing in technology and training, as well as identifying and rewarding employee behaviors that embody a data-led culture. Beyond leadership setting a tone for the whole organization, let's take a look at a few other components that can help build a strong data culture. 1 Bring Business and Data Science Together One of the first steps in building a data culture is to build a strong data science team consisting of data analysts, data engineers, and data scientists. Having quality in-house data talent is a competitive advantage that reaps multiple benefits, including building a robust culture focused on data. Once a data science team is up and running, it needs to be strategically embedded across various departments of the business. This helps business professionals interact with data professionals more regularly and better understand how the power of data analytics and data science can improve business efficiencies and impact profitability and growth. At the same time, this setting enables data professionals to better understand how the business works and build intuition for developing better data and machine learning–powered tools and products. This creates a positive flywheel where both business and data science teams learn to collaborate better and benefit from their respective skill sets. By bringing business and data science together, everyone in the organization learns to appreciate the value of data and use data-driven insights to improve the quality of their decisions, products, and services. 2 Leverage Data When Creating Goals and Deadlines Driving strategic business goals and metrics by leveraging data is a key aspect of encouraging a data-led culture. When goal-setting exercises are conducted objectively and leaders regularly use data and metrics from previous business quarters or external data about competitors or the overall market, everyone in the organization will start to embrace similar data-driven approaches. Leveraging data for setting new targets also enables every stakeholder in the organization to understand and anticipate their future goals and prioritize their work accordingly. Data-led goal setting is a more democratic and fair-minded process that encourages ownership of respective goals by every employee, as opposed to arbitrary, instinct-led, unilateral decisions made by the leadership. 3 Ensure Everybody Has Access to Data A fundamental step toward attaining a data culture is to democratize access to data across the organization. Data culture is a difficult goal when employees in different parts of a business struggle to obtain data. If you don't give your employees access to your data, they won't be able to utilize it when making decisions. This disenfranchises the data analysts, engineers, and scientists disproportionately, as their day-to-day work is impacted the most. Without a motivated team of data professionals, the downstream benefits of data are unlikely to materialize across various business departments. A strong foundation of data governance and data democratization is a prerequisite to achieving the business goals associated with a robust data culture. 4 Keep Your Data Technology Up-to-Date A critical aspect of building a data culture is employing modern tools and technologies to make it easier for employees to access, analyze, and share data-driven insights. Building a modern data stack with newer components like a metrics layer simplifies data-based operations and analytics for everyone, especially nontechnical business stakeholders. Technology, like data warehouses and metrics layers; data analytics tools, like Tableau or Power BI; and customer relationship management (CRM) tools, like Salesforce, are indispensable for modern businesses. Building the data architecture in a cloud environment like Amazon Web Services further improves access to data and reduces the need for multiple tools with a steep learning curve. The right use of tools for data, collaboration, and customer service goes a long way in fostering the use of technology to drive a strong data-led culture. 5 Provide Training for Employees Having supportive leadership and access to data and technology is of little use if employees are not data literate and able to extract insights from data. This requires further investment in terms of learning and development to empower employees with the necessary skills to explore, understand, and share data-driven insights across the organization. In addition to reducing the skills gap, it also encourages people from nontechnical backgrounds to become more data savvy, collaborate better with data experts, and build more comprehensive data products and solutions to benefit the business. 6 Reward Data-Oriented Decisions and Behavior The primary challenge to becoming a data-driven organization is not technical but cultural. A strong data culture is based on a robust foundation of people, policies, and technology. However, once the initial foundation is in place, data leaders need to maintain and bolster the spirit of data-driven decision-making by incentivizing and rewarding behaviors that embody the culture. At the same time, decisions and behaviors that do not represent a holistic data-led process ought to be called out and questioned until every single employee is on board with the philosophy of using data for every decision. This includes encouraging experimentation to answer key business questions for which data does not exist yet or when the current set of data does not yield compelling evidence. Conclusion In this article, you learned about the importance of a data culture for businesses. It's a formidable task to build a strong data culture and is a top priority for a majority of CEOs. Data-driven companies are in a better position to attract and retain talent, make faster decisions with more conviction, and drive stronger growth and profitability to meet their business goals. According to research by McKinsey & Company, data-driven companies are able to achieve their goals faster and realize at least 20 percent more earnings. Related Blogs Introduction
Today, data is at the core of many companies, and it's of the highest importance for running a successful business. Companies process huge amounts of data daily, which they must store, categorize, track, and organize by cataloging, and that's where data governance comes in. Data governance is a set of processes that promote better management of business data, unlocking the true value of data by ensuring that it's more accessible, reliable, secure, and compliant. For modern data-driven organizations, a strong data governance framework is not only important but essential for the best use of data in business decisions. A strong data governance framework usually encompasses functions such as managing data access and data ownership, tracing data lineage, managing duplicate or false data, and classifying and assuring data quality. All of these are the pillars of a successful data governance process. However, implementing a robust data governance framework is no small feat. If not done systematically, it can lead to a huge loss of organizational time, resources, and effort. Companies that have made significant progress in building data governance frameworks and cultivated a strong and inclusive data culture have done so incrementally, aligning incentives and creating deep collaboration across cross-functional teams that own the data governance roadmap. Organizations are more likely to be successful if they can bring together people, processes, and technology to build their framework. In this article, you'll learn about best practices for implementing data governance in an organization. Companies can leverage existing best practices and build on them to fast-track their own data governance efforts. What Are the Challenges of Implementing Data Governance? Before you plan your data governance strategy, you need to look out for some common challenges. One major challenge for organizations is building a strong business use case for investing staff and resources in a data governance framework. Those that haven't yet embraced digital transformation and the better, faster decision-making possible with deeper data analysis might not see the long-term business value of data governance. It's important to unite relevant stakeholders across the organization to take on the challenge. Even when organizations do launch a governance framework, they may fail to achieve its true potential. Poor data leadership and ownership may be an obstacle, for example. Data governance also requires collaboration and consistent enforcement across departments to succeed. For example, the finance department could collaborate with the accountancy department to create a cross-practice team to communicate and transfer data more transparently. So, without the buy-in and blessings of the tech and collaborative data ownership that helps break down the organizational silos, the program is unlikely to come to fruition. Additionally, a good data governance framework relies on high-quality data. The primary goal of data governance is to make data more accessible, secure, and reliable for stakeholders to consume for their own use cases. However, if the quality of the data at the source is poor, implementing data governance becomes much more difficult. Data Governance Best Practices The following are best practices that have been adopted successfully by numerous organizations, such as Collibra, IBM, Informatica, Select Star, and more, in building comprehensive data governance frameworks. 1 Build a Strong Business Use Case The goal of data governance is to enable every stakeholder to use the data to make business decisions relevant to their department, whether that's sales, marketing, finance, or human resources. This means that you need the support and alignment of all users and departments right from the beginning. Without cross-functional support, building a strong business case for investing in a long-term mission like data governance is less likely to succeed. Data governance generates some significant business benefits that can make the advantages of the process clear to the leadership. It saves time and provides improved security and reliable and more accurate data, making it easier to make data-driven decisions. When these business benefits are made clear to the leadership, it's easier to get approval for needed staff, budget, and resources for the project. 2 Identify Data Stewards and Owners Clearly defined roles and owners are necessary to build the data governance framework in a structured manner. Knowing which stakeholders own certain responsibilities also helps with clear lines of communication. Exact roles may differ across organizations, but the following are common choices:
3 Start Small Creating a strong data governance framework requires the right mix of people, processes, and technology to come together. It's crucial to start small and aim for quick incremental wins rather than overpromising and underdelivering. Creating governance guidelines requires specific expertise; you could hire this expertise, but empowering and upskilling people within your existing team might be more successful as they already know your data. Those responsible for data governance then need to gradually build trust and seek alignment from various cross-functional departments before the framework policies can be enshrined as organization-wide processes. For governance-based processes to be adopted and diligently followed, your data stewards need to implement regular checks and audits and guide team members and departments that might not be familiar with good data governance practices. This guidance has two dimensions: cultural guidance and technological guidance concerning the required tools. When data stewards implement processes, they should also implement the right tools for advanced actions such as automation. Once every cross-functional team understands when and how to use governance principles in their day-to-day work with the help of the tools, you can automate some of the processes. 4 Define and Measure Metrics Data governance is a long-term investment. However, it's important to measure progress in smaller time frames to ensure that key milestones are being achieved without any delays or hurdles. Monitoring some metrics, such as the percentage of the data assets per ownership, the number of questions or errors that are reported to the data team, or the number of dashboards that are being used across the organization and their types, might help achieve those key milestones in the long term. In other words, a clear roadmap with specified deliverables, timelines, and metrics that are shared among all the owners ensures that progress can be evaluated in achievable, measurable steps. You need to be able to periodically check the progress of your governance framework to ensure that it's still on track. This image shows a detailed roadmap for establishing a data governance program over a period of two years. Individual tasks can be defined for each business quarter and for different aspects of the framework, such as data insights, data quality, data standards, and data governance and management. For example, improving data quality can be broken down into multiple milestones per business quarter. The goal for the first quarter may be hiring a data engineering team, while the next quarters may focus on establishing reference data repositories, data cleaning, and building data stores and data warehouses. This structured approach keeps cross-functional teams informed on the overall plan and ensures continued progress. 5 Establish Strong Communication Channels Frequent and effective communication is the key to aligning stakeholders and collaborating across teams. Everyone should understand the desired goals and keep others informed on their progress in implementing them. Additionally, your data stewards must be as transparent as possible to earn trust across the organization and emphasize the impact of investment in data governance to the executive leadership as well as to the downstream users of the framework. They can create a single channel for communication, which is like a linked data catalog where you can search data assets or collaborate on them. This way of communication is pivotal both during the implementation phase and after the framework is established. A single channel for communication will help drive strong adoption rates, resolve queries, and allow you to share updates to the governance policies as data and compliance requirements evolve. 6 Contextualize Data Data contextualization involves adding any relevant information to data to make it actionable. Contextualization provides users better interpretation of the data and enables organizations to make smarter decisions. This helps a data governance process work more efficiently as contextualized data has clearer meanings and allows decision makers to have enriched information regarding the actions they should take. Moreover, it can help improve how the organization handles data in its data governance environment. 7 Build a Long-Term Strategy for Data Governance Achieving a strong data governance framework can be a moving target. You need to ensure that stakeholders know this is a long-term investment. Data governance is a continuous process that consists of many smaller projects and deliverables. Ramping up speed and complexity over time helps to scale your efforts. While the overall framework may take several years, smaller milestones can be set and achieved over shorter time frames, like a business quarter. For instance, a useful set of milestones to accomplish in the first quarter of working on a data governance framework may include establishing data management policies and standards, hiring a data engineering team, and drafting a data management strategy together with all relevant stakeholders. As long as they see incremental progress, stakeholders will learn to trust the process and be invested in the success of the project. 8 Expose the Data through Documentation Knowing exactly what your data represents is a critical component of data governance. Users should have a single, centralized platform where they can find documentation related to their data. This documentation should be continuously updated, reviewed, and revised and should also be directly tied to the actual data assets. These actions will ensure that your users can trust and rely on your documentation, as it will always be up to date and accurate. Strong data governance should expose the data through process-oriented documentation that is directly connected to the data. 9 Data Lineage and Usage Knowing the source of data, where your data is flowing, and who is accessing it is important. With data governance, you have to build trust in your data, ensure the data is used properly in your organization, and troubleshoot issues when they arise. Data lineage helps automatically identify sensitive information and propagate some data governance-related policies. Data lineage also informs reports, issue logs, and audit logs, which show that the data governance requirements are met. As an example, data lineage prevents teams from using a dashboard that was supposed to be deprecated or two different business units from building a metric using different underlying data. Successful Data Governance Frameworks Several large global companies have successfully implemented data governance frameworks. The following are some examples. PwC, a global professional services company, has created a data governance framework consisting of the following components:
ING, a Dutch multinational banking and financial services corporation, leveraged IBM Cloud Pak to improve data governance for its users in a hybrid cloud environment. There are also several third-party companies that assist larger organizations with their data governance strategy and implementation, such as Collibra, Informatica, and Alation, and data catalogs that provide tools and insights required for implementing a data governance practice on your own, such as Select Star and Atlan. Outcomes of a Strong Data Governance Implementing a strong data governance strategy will inevitably lead to outcomes such as improved data quality, decreased data management costs, and better data analytics, which, in turn, leads to better decision-making throughout the organization. The following list provides an overview of the outcomes of effective data governance:
For an organization, the time it takes to achieve these outcomes is closely related to the strength of its data governance implementation processes. Over time, these all contribute to one overarching outcome: organizational success. Conclusion Data governance is an essential requirement for modern organizations to drive greater adoption of data and empower business decision-making. Organizations can find it difficult to extract the full value of their data assets, especially as the amount of data keeps growing. Data governance frameworks lay down clear policies and guidelines for improving the quality of data and democratizing its usage across a business. If you can navigate the challenges involved and follow the above best practices in creating and implementing your data governance framework, you can accelerate your organization's understanding and usage of data and deliver data-driven decision-making to your organization. Related Blogs
Published by Andela Introduction
Modern tech companies realize that data teams need to consist of professionals with varied expertise, including data analysts, data engineers, data scientists, applied scientists, and machine learning engineers. Data teams work closely with cross-functional stakeholders to build data-driven products that are powered by predictive analytics as well as machine learning. Data-driven organizations rely on robust data infrastructure and ETL processes for downstream machine learning use cases. This recent development is accompanied by the rise of data engineering as a specialized discipline. As more organizations undergo digital and AI transformation journeys, the demand for data engineers has increased concomitantly. Data engineers are required to build the data infrastructure and pipelines and facilitate easy access to processed data for data scientists to build machine learning models. In this article, we’ll dive into the differences between the profiles of a data engineer and a data scientist along several dimensions, including their roles and responsibilities, educational requirements, specializations, and career growth. Roles and responsibilities of data engineers and data scientists Data engineers primarily build the pipeline system for data scientists to consume with models for various use cases. Therefore, data engineers are often hired earlier to build the data platform before onboarding data scientists. In smaller companies and startups, it is not uncommon for data professionals to do both data engineering and data science. As a company grows and scales its data science efforts, specialized data engineering and data science professionals become necessary. Data engineer’s responsibilities
Data scientist’s responsibilities
Every day, data engineers usually write code, build data pipelines, and maintain various pieces of the data infrastructure as well as serve requests for cleaned and processed data from data scientists. Data scientists typically spend most of the day developing and training machine learning models, conducting multiple experiments to optimize the model performance, and meeting cross-functional stakeholders from engineering, product, and business teams to discuss results and develop new use cases. Education differences between data engineers and data scientists Data engineers typically have a bachelor’s degree in computer science or information technology. Their core expertise is focused on software engineering skills such as programming, algorithms, data structures, systems architecture, and building software tools. With the advent of cloud computing as the foundation for any tech organization, data engineers are also expected to be familiar with relevant cloud-based technologies (like AWS, Microsoft Azure, and Google Cloud Platform) focused on data warehousing, data visualization, and data analytics. Similarly, data scientists are also able to leverage cloud-based machine learning services and APIs for common use cases such as recommender systems, computer vision, and NLP, instead of starting from scratch. Certifications provided by these cloud companies are often mandated as compulsory training during the onboarding phase for new data scientist and data engineer candidates. As data engineering is focused on building data systems for data scientists, engineers require a better understanding of statistics or machine learning to help communicate and collaborate with the rest of the data team. Data scientists have a more diverse background with undergraduate-level training in computer science, statistics, mathematics, physics, psychology, and life sciences. Data scientists often have more advanced degrees, such as a master’s degree or a PhD, in any of the above disciplines. Though data scientists traditionally had more advanced degrees, particularly the first wave which emerged a decade ago, it is becoming increasingly common for entry-level data science jobs to not have such requirements. Additionally, data scientists work with multiple stakeholders from engineering, analytics, product, and business teams, and it is helpful for them to know a bit about these areas for a smoother and more efficient collaboration. Building a successful, collaborative data product with diverse cross-functional teams requires efficient communication and storytelling skills from data scientists. Specializations With the rising popularity of data science and data engineering jobs, a number of upskilling platforms, courses, and boot camps now offer specialized, practical, hands-on training. These specializations are industry oriented and often developed by leading tech companies such as Google, Microsoft, AWS, IBM, etc. There are also many certification courses that allow candidates to learn specific data skills and signal their motivation and skill set to prospective employers. The following are a selection of specializations or certifications that a successful data engineer may have:
The following are a selection of specializations or certifications that a successful data scientist may have:
However, prospective data engineers or scientists must carefully consider which course is best suited to them given the constraints of finances, time, and interests. It is not feasible nor necessary to undertake as many courses as possible, and it is more important to focus on the courses that can truly improve your understanding and improve your candidature as a data engineer or a data scientist. Career growth differences between data engineers and data scientists Career growth prospects for both data engineers and data scientists are promising. Data engineers can evolve into related roles such as data architect or solutions architect. They can become leaders who envision and lead teams working on data platforms and also transition into more traditional engineering leadership roles. With a better understanding of core data science skills such as statistics and machine learning, data engineers can also switch to data scientist roles. The demand for data scientists has remained consistently strong for over a decade now. There are numerous entry-level positions at companies of all sizes and business domains. Initially restricted to experts with deep domain expertise and doctoral training, data science has now become more democratic with the development of tools and technologies that simplify and automate the various nuts and bolts of the data science lifecycle. Data scientists can progress further to become recognized domain experts as individual contributors or build data science teams and organizations as data science leaders. With a better grasp of software engineering fundamentals such as data structures, algorithms, and optimized coding, data scientists can also switch laterally to become data engineers or machine learning engineers. Final thoughts With rapid advances in data science and the increasing appreciation for its value in business growth, companies are actively building their data science teams and capabilities. The first step involves building the foundational infrastructure for data, a job that is carried out by data engineers. They take care of building data warehouses and pipelines and provide data that is ready to be consumed by data scientists for building various machine learning models and applications. Related Blogs
Introduction
Consumer technology companies like Amazon, Yelp, and Airbnb are focused on providing an impeccable customer experience, and reviews are integral to that experience. Reviews from previous customers can signal trust and reliability (e.g., total number of reviews or average star rating), empowering first-time buyers or new customers in their decision-making. Millions of reviews are shared on platforms like Amazon for e-commerce products, on Airbnb for travel and hospitality, on Glassdoor for company and employment experience, and on Google for third-party businesses. However, the internet has become rife with fake reviews. Fake reviews and inflated ratings provide a tainted picture of a product or service and are designed to trick customers away from or toward certain purchases. As these reviews are an important input factor for search and ranking algorithms, they can have a massive influence on product discovery and sales. This provides a strong incentive for bad actors to try to manipulate the system by improving the ratings of their products through fake reviews. There is a booming market for fake reviews, which are purchased via multiple social media and community platforms. The problem is enormous - nearly four percent of all reviews are fake, translating into a global economic impact of USD 152 billion. E-commerce companies like Amazon spend upwards of a billion dollars and employ tens of thousands of workers to combat online fraud and abuse. Some companies use sophisticated technologies including AI to detect and delete fake reviews, but their accuracy is limited (less than forty percent) and it often takes more than one hundred days to remove those reviews. During that time, fraudulent sellers can make strong short-term revenues and profits. Apart from the short-term commercial losses, there is a longer-term problem; fake reviews erode customer trust and safety, causing customers to avoid online purchases. Catching fake reviews is therefore paramount for a majority of online marketplaces and businesses. Characteristics of Fake Reviews Fake reviews have several telltale characteristics. For instance, as they are based on a fraudulent experience with the product or service, fake reviews will often focus on a poor customer experience without specific details about that product or service. Another sign is the repetition of positive or negative keywords and text. As it is difficult to fabricate a review, fake reviewers keep emphasizing certain keywords and details to paint a terrible customer experience. Such reviews accentuate extreme details without providing a balanced perspective. Fake reviewers also excessively use emoticons and exclamation points in an attempt to appeal to the customers’ emotions. Genuine reviewers tend to focus more on information and provide thoughtful, actionable feedback for other customers about the product experience. One clear giveaway is the reviewer’s name and avatar. Fake reviews are usually submitted using an account with a dubious username, avatar, or email address. If a reviewer seems like they could be illegitimate, check whether they have shared any reviews previously, how often, and for which products or businesses. Fake accounts are often created for one-time use, and fake reviewers can submit multiple reviews in a short span of time, sometimes on the same day. Sometimes fake reviewers post a poor rating without any comments to describe their experience. Genuine reviewers take the time and effort to write useful feedback. Spotting these characteristics can help you find many fake reviews, but scamsters are always devising more sophisticated techniques to replace those that have already been detected through algorithms, AI, or human reviewers. One solution to this problem is fingerprinting technology, that can identify unique users of your website regardless of VPNs, cookie blockers, private browsing, or other tools. They use data including the browser and device used, usage patterns, IP addresses, and geolocation to create a unique identifier for site visitors, making it easier to spot users trying to hide their identity or committing fraudulent activity. Conclusion Fake reviews have undermined the revenue and growth of online sellers and small businesses. These reviews can boost the sales of a poor product by exaggerating its positive rating, or damage the sales of competitor products via negative reviews. While there are ways to catch these fake reviewers in the act, it’s an increasingly sophisticated scam and a headache for businesses. Fingerprinting technology can help you find and remove fake reviews as well as protect your business from all types of online fraud. This helps ensure that your customers will have a safe and reliable online shopping experience. Recently, the Government of India issued a draft framework of standards to counter fake reviews in order to reduce their prevalence on e-commerce platforms. Related Blog Published by CloudForecast Introduction
Amazon Redshift is a widely used cloud data warehouse that is used by many businesses, like Nasdaq, GE, and Zynga, to process analytical queries and analyze exabytes of data across databases, data lakes, data warehouses, and third-party data sets. There are multiple use cases for Redshift, including enhancing business intelligence capabilities, increasing developer and analyst productivity, and building machine learning models for predictive insights, like demand forecasting. Amazon Redshift can be leveraged by modern data-driven organizations to vastly improve their data warehousing and analytics capabilities. However, the pricing for Redshift services can be challenging to understand, with multiple criteria that define the total cost. In this article, you’ll learn about Amazon Redshift and its pricing structure, with suggestions for how to optimize costs. What Is Amazon Redshift? Essentially, Amazon Redshift provides analytics over multiple databases and offers high scalability in a secure and compliant fashion. Additionally, there is a serverless option called Amazon Redshift Serverless that makes it even easier to rapidly scale analytics setup without requiring a managed data warehouse infrastructure. It helps with data democratization and assists various data stakeholders to extract data insights by simply loading and querying data in the warehouse. Amazon Redshift Pricing In this section, you’ll learn about Amazon Redshift’s capabilities as it pertains to usage and pricing. Free Tier For new enterprise users, the AWS Free Tier provides a free two-month trial of the DC2.Large node. This free service includes 750 hours per month, which is sufficient to run a single DC2.Large node with 160GB of compressed solid-state drives (SSD). On-Demand Pricing When you launch an Amazon Redshift cluster, you select a number of nodes in a specific region as well as their instance type to run your data warehouse. In on-demand pricing, a simple hourly rate applies based on the previous configuration and is billed as long as the cluster is live. The typical hourly rate for a DC2.Large node is $0.25 USD per hour. Redshift Serverless Pricing With Amazon Redshift Serverless, costs accrue only when the data warehouse is active and is measured in units of Redshift Processing Units (RPUs). You’re charged in terms of RPU-hours on a per-second basis. The serverless configuration also includes concurrency scaling and Amazon Redshift Spectrum, and the cost for these services is already included. Managed Storage Pricing Amazon Redshift charges for the data stored in a managed storage at a specific rate per GB-month. Its usage is calculated on an hourly basis as a function of the total amount of data and starts as low as $0.024 USD per GB with the RA3 node. The cost of a managed storage also varies according to the particular AWS region in which the data is stored. For example, consider the cost of a managed storage pricing where 100TB of data is stored with an RA3 node type for thirty days in the US East region, where the cost is $0.024 USD per GB-month. The total usage for thirty days in GB-hours is as follows: 100TB × 1024GB/TB (converting TB to GB) × 30 days × 24 hours/day = 73,728,000 GB-hours Then you can convert GB-hours to GB-months: 73,728,000 GB-hours / (24 × 30) hours per month = 102,400 GB-months Finally, you can calculate the total cost of 102,400 GB-months at $0.024 USD/GB-month in the US East region: 102,400 GB-months × $0.024 USD = $2,457.60 USD Spectrum Pricing With Amazon Redshift Spectrum, users can run SQL queries directly on the data in the S3 buckets. Here, the cost is based on the number of bytes scanned by the Spectrum utility. The pricing of Redshift Spectrum is $5 USD per terabyte of data scanned. Concurrency Scaling Pricing With Concurrency Scaling, Amazon Redshift can be scaled to multiple concurrent users and queries. For every twenty-four hours that your main cluster is live, you accrue a one-hour credit. Any additional usage is charged on a per-second, on-demand rate that depends on the number of types of nodes in the main cluster. Reserved Instance Pricing Reserved instances are designated for stable production workloads and are less expensive than clusters run on an on-demand basis. Significant cost savings can be achieved through long-term usage and commitment to Amazon Redshift in the span of a few years. Pricing for reserved instances can either be paid all up front, partially up front, or monthly over the course of a year with no up-front charges. Amazon Redshift Cost Optimization Considerations Before you begin using Amazon Redshift, you need to be aware of your current costs. AWS Cost ExplorerThe AWS Pricing Calculator provides a configurable tool to estimate the cost of using Amazon Redshift. For instance, the annual cost of one node of the DC2.8xlarge instance in the US East (Ohio) region on an on-demand basis is as follows: 1 instance × $4.80 USD hourly × 730 hours in a month × 12 months = $42,048 USD The cost for the same Amazon Redshift configuration for a reserved instance for a one-year term paid up front is $27,640 USD. AWS Tags Using AWS cost allocation tags can help you decode and manage your AWS costs. Tagsenable AWS resources to be labeled in the form of key-value pairs and can include various types, like technical, business, security, and automation. Once the tags are activated in the Billing and Cost Management console, a cost allocation report can be generated based on the specific resources tagged. Tags can be user-defined or AWS-generated. Amazon Redshift Cost Optimization Optimizing Amazon Redshift costs comes down to effective planning, prudent usage and allocation of resources, and regular monitoring of the usage and associated costs. Optimizing Queries The analytical queries made on the data stored in Amazon Redshift can be optimized to run more efficiently. Queries can be compute-intensive, can be storage-intensive, or can take a long time to execute. There are a number of query tuning techniques that can be used to optimize your queries. Tables with skewed data or missing statistics, and queries with nested loops and long wait times, typically affect query performance and can be improved as illustrated in this AWS developer guide. Here is a commonly used weak query that selects all the columns in a table: SELECT * FROM USERS The previous query can be very inefficient and slow if the table consists of thousands of columns, especially if only a few columns are relevant for the necessary analysis. This query can be optimized by specifying and retrieving the exact column names like the following: SELECT Firstname, Lastname, DOB FROM USERS Cluster Limits and Quotas Usage limits on Amazon Redshift clusters can be programmed using the AWS Command Line Interface (CLI) tool. Limits can be imposed on concurrency scaling in terms of time and spectrum in terms of data scanned. Daily, weekly, or monthly periods can be used. A number of limits and quotas are defined for Redshift resources that can also be applied to constrain the overall costs associated with Redshift. Data Type Amazon Redshift costs can also be managed by storing data in a compressed, partitioned, and columnar data format, like Apache Parquet, since fewer data is scanned. Conclusion Amazon Redshift is a powerful and cost-effective cloud-native data warehouse that provides scalable and performant data analytics and processing capabilities. It also comes with a serverless configuration that allows any data stakeholder to run data queries without the need to provision and manage the data warehouse infrastructure. Amazon Redshift has multiple aspects affecting its pricing, including on-demand or reserved capabilities, serverless, managed storage pricing, Redshift Spectrum pricing, concurrency scaling pricing, and reserved instance pricing. Keeping on top of the various Amazon Redshift costs is not straightforward but can be made easier by AWS cost monitoring tools, like CloudForecast. CloudForecast helps manage AWS costs through daily cost management reports, monthly financial reports, untagged AWS resources discovery, and idle and underutilized resources visibility for cost-saving opportunities. Related blog Published by CloudForecast Introduction
Companies are increasingly moving their production code to serverless functions using AWS Lambda, which has gained popularity for its better code maintenance, low-cost hosting charges, and automatically scaled and optimized performance. But without careful oversight, Lambda can become an expensive choice for your project. Lambda, offered by market-leading AWS, offers many benefits. Lambda is one example of serverless functions, or single-purpose, programmatic functions hosted and maintained by cloud providers like AWS, Azure, or GCP to ensure near-perfect runtime and scaling to any incoming network request volume. Companies can use Lambda, an event-driven compute service, to run any type of application or backend service without worrying about provisioning or managing servers. Lambda adapts to a variety of use cases across startups and enterprises alike. It can process data at scale, run interactive web and mobile backend services, enable powerful machine learning models, and build in-house event-driven applications. It also specifies limits for the amount of compute and storage resources used to run and store serverless functions. These limits apply to a number of resources, such as the number of concurrent executions; storage for uploaded functions as well as quotas for function configuration; deployment and execution parameters like memory allocation; timeout; environment variables; layers; and burst concurrency. The key to using Lambda is keeping your costs in check. This article will review Lambda’s pricing structure to show how costs can be efficiently managed without compromising on operational excellence and execution of Lambda functions. It will also discuss tools like CloudForecast that can help engineering teams monitor and reduce their serverless computing costs on AWS. Understanding AWS Lambda Pricing AWS Lambda pricing is based on the amount of memory allocated to the serverless function and the amount of time the code runs, rounded to the nearest millisecond. The key variables that determine Lambda costs are the type of architecture, the number of requests, the time frame for which the requests apply, the duration of each request (in milliseconds), and the amount of memory allocated to the Lambda function. Each Lambda request starts when code executes in response to an event trigger from services like Amazon’s Simple Notification Service or calls from Amazon API Gateway or via the AWS SDK. The cost for each compute and storage resource is calculated depending on the function configuration. AWS offers a free tier that allows one million free requests per month and 400,000 GB-seconds of compute time per month powered by x86 and Graviton2 processors. It also offers a flexible pricing model called the Compute Savings Plan, based on guaranteed usage (measured in dollars per hour) for a one- or three-year term. AWS Lambda does offer an attractive feature called Provisioned Concurrency that enables greater control over start-up latency when Lambda functions are triggered. Provisioned concurrency solves the problem of variable start-up latency when a Lambda service is triggered on demand and scales up to meet the needs of the application workloads. This overhead in starting a Lambda function is referred to as cold start, and the magnitude of this problem is a function of the time taken to set up the execution environment and the duration for the code to be initialized. As illustrated in this official AWS example, with provisioned concurrency enabled, the percentage of requests served within a given time remains fairly constant—especially for the slowest five percent of the requests—in comparison to a scenario with provisioned concurrency disabled. At scale, this can have a massive impact not only on the costs but also on the user experience. While the first factor is controlled by AWS, the second factor falls to the developer. The code initialization duration is predominantly responsible for cold start latency. Provisioned concurrency solves for cold start by enabling Lambda functions to be initialized for high workloads in milliseconds. AWS provides a pricing calculator to estimate the cost of using Lambda for your applications. The below estimate provides pricing calculations for a sample application with the following settings:
The same pricing calculator can also provide an estimate for provisioned concurrency. In this case, in addition to the above parameters, the cost is a function of the amount of concurrency specified and the period of time the configuration is active. Controlling AWS Lambda Costs AWS Lambda does offer options for controlling costs, but as the above example showed, the cost of function calls can quickly scale up as part of the organizational application workload. If the configuration is not carefully monitored and fine-tuned for current applications, Lambda can become prohibitively expensive. You can keep AWS Lambda costs down by focusing on three important factors:
The cost of a Lambda function invocation is multiplied by its execution time and memory size, so reducing either factor by even a small amount can have a significant impact on billable costs. It’s important to ensure you have the correct configuration. Periodic monitoring of the actual values of the memory size and the number and duration of function calls can help confirm whether the current configuration is fine-tuned for the current workload. AWS Lambda logs are ingested into Amazon CloudWatch, so mining these logs can help optimize the configuration and the costs. External tools like CloudForecast can also monitor usage and costs. Avoiding high maximum execution time also helps save costs. It’s common to have a buffer of execution time beyond what’s specified, but the additional costs incurred by Lambda functions add up, making it prudent to change the value of the “duration of each request” parameter as needed. Lambda Step Functions can also help manage costs. Step functions are state machines with a visual workflow that allow developers to coordinate different tasks like calling various Lambda functions. Using step functions is a more efficient way to poll for the status of tasks. Typically, long polling increases the costs of Lambda functions as they are waiting idle, and step functions help alleviate the total costs based on the number of state transitions to execute the application, instead of the execution time of a workflow. Another tactical method to control Lambda costs is to evaluate whether your application can be run asynchronously. Running async workloads prevents idle downtime in which the AWS Lambda functions wait for external applications to complete. If the overall architecture can be analyzed for idle instances and reconfigured for asynchronous execution, the costs of Lambda functions can be drastically reduced. The frequency at which Lambda functions are invoked can also impact the usage and costs. Where applications like Kinesis are used as a Lambda function trigger, increasing the batch size can reduce the frequency at which the Lambda function needs to be invoked, thus reducing the total number of executions. Writing optimized production code always helps, and its lower execution time can reduce Lambda costs. You can, for instance, record and analyze the Duration metric in CloudWatch for slow execution times. For some applications, EC2 spot instances may be cheaper and more effective than Lambda functions. This is especially true for an application architecture in which the traffic is predictable and sustained, making a reliable EC2 spot instance a more suitable alternative. Conclusion AWS Lambda and serverless functions have had a tremendous impact on the efficient execution of software, data, and machine learning applications in the cloud. Lambda can help you achieve savings on your engineering costs, but it’s possible to reduce your costs even more by optimizing the configuration of your applications and fine-tuning your resources. Doing this work manually can require careful logging and monitoring of your application in production settings. Instead, you can use tools to automate and dynamically adjust Lambda function settings to reduce costs in a more cost- and time-efficient manner. One of those tools is CloudForecast, which can manage and optimize the cost of using AWS services like Lambda. CloudForecast provides an out-of-the-box solution for engineering teams to monitor their monthly budget and move toward a more responsible use of Lambda functions. Its detailed reports suggest ways to reduce AWS costs, and it can also provide reports for your finance and accounting teams. To learn more about how CloudForecast can help with your AWS Lambda costs, check its official blog. Related blog Data science teams are an integral part of early-stage or growth-stage start-ups as midlevel and enterprise companies. A data science team can include a wide range of roles that take care of the end-to-end machine learning lifecycle from project conceptualization to execution, delivery, and monitoring:
The manager of a data science team in an enterprise organization has multiple responsibilities, including the following:
As the data science manager, it’s critical to have a structured, efficient hiring process, especially in a highly competitive job market where the demand outstrips the supply of data science and machine learning talent. A transparent, thoughtful, and open hiring process sends a strong signal to prospective candidates about the intent and culture of both the data science team and the company, and can make your company a stronger choice when the candidates are selecting an offer. In this blog, you’ll learn about key aspects of the process of hiring a top-class data science team. You’ll dive into the process of recruitment, interviewing, and evaluating candidates to learn how to find the ones who can help your business improve its data science capabilities. Benefits of an Efficient Hiring Process Recent events have accelerated organizations’ focus on digital and AI transformation, resulting in a very tight labor market when you’re looking for data sciencedigital skills, like machinelike data science and machine learning, statistics, and programming. A structured, efficient hiring process enables teams to move faster, make better decisions, and ensure a good experience for the candidates. Even if candidates don’t get an offer, a positive experience interacting with the data science and the recruitment teams makes them more likely to share good feedback on platforms like Glassdoor, which might encourage others to interview at the company. Hiring Data Science Teams A good hiring process is a multistep process, and in this section, you’ll look at every step of the process in detail. Building a Funnel for Talent Depending on the size of the data science team, the hiring manager may have to assume the responsibility of reaching out to candidates and building a pipeline of talent. In larger organizations, managers can work with in-house recruiters or even third-party recruitment agencies to source talent. It’s important for the data science managers to clearly convey the requirements for the recruited candidates, such as the number of candidates desired and the profiles of those candidates. Candidate profiles might include things like previous experience, education or certifications, skill set or tech stack, and experience with specific use cases. Using these details, recruiters can then start their marketing, advertising, and outreach campaigns on platforms, like LinkedIn, Glassdoor, Twitter, HackerRank, and LeetCode. In several cases, recruiters may identify candidates who are a strong fit but who may not be on the job market or are not actively looking for new roles. A database of all such candidates ought to be maintained so that recruiters can proactively reach out to them at a more suitable time and reengage the candidates. Another trusted source of identifying good candidates is through employee referrals. An in-house employee referral program that incentivizes current employees to refer candidates from their network is often an effective way to attract the specific types of talent you’re looking for. The data science leader should also publicize their team’s work through channels, like conferences or workshops, company blogs, podcasts, media, and social media. By investing dedicated time and energy in building up the profile of the data science team, it’s more likely that candidates will reach out to your company seeking data science opportunities. When looking for a diverse set of talent, the search an be difficult as data science is a male dominated field. As a result, traditional recruiting paths will continue to reflect this bias. Reaching out and building relationships with groups such as Women in Data Science, can help broad the pipeline of talent you attract. Defining Roles and Responsibilities Good candidates are more likely to apply for roles that have a clear job description, including a list of potential data science use cases, a list of required skills and tech stack, and a summary of the day-to-day work, as well as insights into the interviewing process and time lines. Crafting specific, accurate job descriptions is a critical—if often overlooked—aspect of attracting candidates. The more information and clarity you provide up front, the more likely it is that candidates have sufficient information to decide if it’s a suitable role for them and if they should go ahead with the application or not. If you’re struggling with creating this, you can start with an existing job description template and then customize it in accordance with the needs of the team and company. It's also critical to not over populate a job description with every possible skill or experience you hope a candidate brings. That will narrow your potential applicant pool. Instead focus on those skills and experiences that are absolutely critical. The right candidate will be able to pick up other skills on the job. It can be useful for the job description to include links to any recent publications, blogs, or interviews by members of the data science team. These links provide additional details about the type of work your team does and also offer candidates a glimpse of other team members. Here are some job description templates for the different roles in a data science team: Interviewing process When compared to software engineering interviews, the interview process for data science roles is still very unstructured, and data science candidates are often uncertain about what the interview process involves. The professional position of data scientist has only existed for a little over a decade, and in that time, the role has evolved and transformed, resulting in even newer, more specialized roles, such as data engineer, machine learning engineer, applied scientist, research scientist, and product data scientist. Because of the diversity of roles that could be considered data science, it’s important for a data science manager to customize the interviewing process depending on the specific profile they’re seeking. Data scientists need to have expertise in multiple domains, and one or more second-round interviews can be tailored around these core skills:
Given how tight the job market is for data science talent, it’s important to not over complicate the process. The more steps in the process, the longer it will take and the higher the likelihood you will lose viable candidates to other offers. So be thoughtful in your approach and evaluate it periodically to align with the market. Types of Data Science Interviews Interviews are often a multistep process and can involve multiple steps of assessments. Screening Interviews To save time, one or more screening rounds can be conducted before inviting candidates for second-round interviews. These screening interviews can take place virtually and involve an assessment of essential skills, like programming and machine learning, along with a deep dive into the candidate’s experience, projects, career trajectory, and motivation to join the company. These screening rounds can be conducted by the data science team itself or outsourced to other companies, like HackerRank, HackerEarth, Triplebyte, or Karat. Onsite Interviews Once candidates have passed the screening interviews, the top candidates will be invited to a second interview, either virtually or in person. The data science manager has to take the lead in terms of coordinating with internal interviewers to confirm the schedule for the series of interviews that will assess the candidate’s skills, as described earlier. On the day of the second-round interviews, the hiring manager needs to help the candidate feel welcome and explain how the day will proceed. Some companies like to invite candidates to lunch with other team members, which breaks the ice by allowing the candidate to interact with potential team members in a social setting. Each interview in the series should start by having the interviewer introduce themself and provide a brief summary of the kind of work they do. Depending on the types of interviews and assessments the candidate has already been through, the rest of the interview could focus on the core skill set to be evaluated or other critical considerations. Wherever possible, interviewers should offer the candidate hints if they get stuck and otherwise try to make them feel comfortable with the process. The last five to ten minutes of each interview should be reserved for the candidate to ask questions to the interviewer. This is a critical component of second-round interviews, as the types of questions a candidate asks offer a great deal of information about how carefully they’ve considered the role. Before the candidate leaves, it’s important for the recruiter and hiring manager to touch base with the candidate again, inquire about their interview experience, and share time lines for the final decision. Technical Assessment It is common for there to be some sort of case study or technical assessment to get a better understanding of a candidate’s approach to problem solving, dealing with ambiguity and practical skills. This provides the company with good information about how the candidate may perform in the role It also is an opportunity to show the candidate what type of data and problems they may work on when working for you. Evaluating candidates After the second-round interviews and technical assessment, the hiring manager needs to coordinate a debrief session. In this meeting, every interviewer shares their views based on their experience with the candidate and offers a recommendation if the candidate should be hired or not. After obtaining the feedback from each member of the interview panel, the hiring manager also shares their opinion. If the candidate unanimously receives a strong hire or a strong no-hire signal, then the hiring manager’s decision is simple. However, there may be candidates who perform well in some interviews but not so well in others, and who elicit mixed feedback from the interview panel. In cases like this, the hiring manager has to make a judgment call on whether that particular candidate should be hired or not. In some cases, an offer may be extended if a candidate didn’t do well in one or more interviews but the panel is confident that the candidate can learn and upskill on the job, and is a good fit for the team and the company. If multiple candidates have interviewed for the same role, then a relative assessment of the different candidates should be considered, and the strongest candidate or candidates, depending on the number of roles to be filled, should be considered. While most of the interviews focus on technical data science skills, it’s also important for interviewers to use their time with the candidate to assess soft skills, like communication, clarity of thought, problem-solving ability, business sense, and leadership values. Many large companies place a very strong emphasis on behavioral interviews, and poor performance in this interview can lead to a rejection, even if the candidate did well on the technical assessments. Job Offer After the debrief session, the data science manager needs to make their final decision and share the outcome, along with a compensation budget, with the recruiter. If there’s no recruiter involved, the manager can move directly to making the candidate an offer. It’s important to move quickly when it comes to making and conveying the decision, especially if candidates are interviewing at multiple companies. Being fast and flexible in the hiring process gives companies an edge that candidates appreciate and take into consideration in their decision-making process. Once the offer and details of compensation have been sent to the candidate, it’s essential to close the offer quickly to prevent candidates from using your offer as leverage at other companies. Including a deadline for the offer can sometimes work to the company’s advantage by incentivizing candidates to make their decision faster. If negotiations stretch and the candidate seems to lose interest in the process, the hiring manager should assess whether the candidate is really motivated to be part of the team. Sometimes, it may move things along if the hiring manager steps in and has another brief call with the candidate to help remove any doubts about the type of work and projects. However, additional pressure on the candidates can often work to your disadvantage and may put off a skilled and motivated candidate in whom the company has already invested a lot of time and money. Conclusion In this article, you’ve looked at an overview of the process of hiring a data science team, including the roles and skills you might be hiring for, the interview process, and how to evaluate and make decisions about candidates. In a highly competitive data science job market, having a robust pipeline of talent, and a fast, fair, and structured hiring process can give companies a competitive edge. Related Blogs Published by Domino Data Lab Reproducibility is a cornerstone of the scientific method and ensures that tests and experiments can be reproduced by different teams using the same method. In the context of data science, reproducibility means that everything needed to recreate the model and its results such as data, tools, libraries, frameworks, programming languages and operating systems, have been captured, so with little effort the identical results are produced regardless of how much time has passed since the original project.
Reproducibility is critical for many aspects of data science including regulatory compliance, auditing, and validation. It also helps data science teams be more productive, collaborate better with nontechnical stakeholders, and promote transparency and trust in machine learning products and services. In this article, you’ll learn about the benefits of reproducible data science and how to ingrain reproducibility in every data science project. You’ll also learn how to cultivate an organizational culture that promotes greater reproducibility, accountability, and scalability. What does it mean to be reproducible? Machine learning systems are complex, incorporating code, data sets, models, hyperparameters, pipelines, third-party packages, model training and development configurations across machines, operating systems, and environments. To put it simply, reproducing a data science experiment is difficult if not impossible if you can’t recreate the exact same conditions used to build the model. To do that, all artifacts have to be captured and versioned in an accessible repository. That way when a model needs to be reproduced, the exact environment, using the exact training data and code, within the exact package combination can be recreated easily. Too often it's an archeological expedition that can take weeks or months (or potentially never) when the artifacts are not captured at the time of creation. While the focus on reproducibility is a phenomenon in data science, it has been a cornerstone of scientific research across all kinds of industries, including clinical and life sciences, healthcare, and finance. If your company is unable to produce consistent experimental results, that can significantly impact your productivity, waste valuable resources, and impair decision-making. Situations Where Reproducibility Matters In data science, reproducibility is especially vital for data scientists to apply the experimental findings to their own work. Regulatory Compliance In highly regulated industries like insurance, finance and life sciences, all aspects of a model have to be documented and captured to provide full transparency, justification and validation on how models are developed and used inside an organization. This includes the type of algorithm being used, why the algorithm has been selected and how the model has been implemented within the business. A big part of complying involves being able to exactly reproduce the results of a model at any time. Without a system for capturing the artifacts, code, data, environment, packages and tools used to build a model this can be a time consuming, difficult task. Model Validation In all industries models should be validated prior to deployment to ensure the results are repeatable, understood and the model will achieve its intended purpose. Too often this is a time intensive process with validation teams having to piece together the environment, tools, data and other artifacts that were used to create the model, which slows down moving a model into production. When an organization is able to reproduce a model instantly, validators can focus on their core function of ensuring the model is robust and accurate. Collaboration Data science innovation happens when teams are able to collaborate and compound knowledge. It doesn’t happen when they have to spend time painstakingly recreating a prior experiment or accidentally duplicate work. When all work is easily reproducible, and easily searched, it's easy to build on prior work to innovate. It also means that as team staffing changes, institutional knowledge doesn’t disappear. Ingraining Reproducibility in Data Science Projects Instilling a culture of reproducibility in data science across an organization requires a long-term strategy, technology investment, and buy-in from data and engineering leadership. In this section, you’ll learn about a few established best practices for conducting and promoting reproducible data science work in your industry. Version Control Version control refers to the process of tracking and managing changes to artifacts, like code, data, labels, models, hyperparameters, experiments, dependencies, documentation, as well as environments for training and inference. The building blocks of version control for data science are more complex than software projects, making reproducibility that much more difficult and challenging. For code, there are multiple platforms, like GitHub, GitLab, and Bitbucket, that can be used to store, update, and track code, like Python scripts, Jupyter Notebooks, and configuration files, in common repositories. However that isn’t sufficient. Datasets need to be captured and versioned as well. So do the environments, tools and packages. This is because code may or may not run the same on a different version of Python or R, for example. Data may have changed even if pulled with the same parameters. Similarly capturing different versions of models and corresponding hyperparameters for each experiment is important to reproduce and replicate the results of a winning model that might be deployed to production. Reproducing end-to-end data science experiments is a complex, technical challenge that can be achieved much more efficiently using platforms like Domino’s Enterprise MLOps platform which eliminates all manual work and ensures reproducibility at scale. Scalable Systems Building accurate and reproducible data science models requires robust and scalable infrastructure for data storage and warehousing, data pipelines, feature stores, model stores, deployment pipelines, and experiment tracking. For machine learning models that serve predictions in real time, the importance of reproducibility is even higher in order to quickly resolve bugs and performance issues. End-to-end machine learning pipelines involve multiple components, and an organizational strategy for reproducible data science work must carefully plan for the tooling and infrastructure to enable it. Engineering reproducible workflows requires sophisticated tooling to encompass code, data, models, dependencies, experiments, pipelines, and runtime environments. For many organizations, it makes sense to buy (vs. build) such scalable workflows focused on reproducible data science. Conclusion Reproducible research is a cornerstone of scientific research. Reproducibility is especially significant for cross-functional disciplines like data science that involve multiple artifacts, like code, data, models, and hyperparameters, as well as a diverse set of practitioners and stakeholders. Reproducing complex experiments and results is, therefore, essential for teams and organizations when making important decisions like which models to deploy, identifying root causes when the models break down, and building trust in data science work. Reproducing data science results requires a complex set of processes and infrastructure that is not easy or necessary for many teams and companies to build in-house. Related Blogs Published by Unbox.ai Introduction
Machine learning models, especially deep neural networks, are trained using large amounts of data. However, for many machine learning use cases, real-world data sets do not exist or are prohibitively costly to buy and label. In such scenarios, synthetic data represents an appealing, less expensive, and scalable solution. Additionally, several real-world machine learning problems suffer from class imbalance—that is, where the distribution of the categories of data is skewed, resulting in disproportionately fewer observations for one or more categories. Synthetic data can be used in such situations to balance out the underrepresented data and train models that generalize well in real-world settings. Synthetic data is now increasingly used for various applications, such as computer vision, image recognition, speech recognition, and time-series data, among others. In this article, you will learn about synthetic data, its benefits, and how it is generated for different use cases. What is synthetic data? Synthetic data is a form of data augmentation that is commonly used to address overfitting deep learning models. It’s generated with algorithms as well as machine learning models to have similar statistical properties as the real-world data sets. For data-hungry deep learning models, the availability of large training data sets is a massive bottleneck that can often be solved with synthetic data. Additionally, synthetic data can be used for myriad business problems where real-world data sets are missing or underrepresented. Several industries—like consumer tech, finance, healthcare, manufacturing, security, automotive, and robotics—are already benefiting from the use of synthetic data. It helps avoid the key bottleneck in the machine learning lifecycle of the unavailability of data and allows teams to continue developing and iterating on innovative data products. For example, building products related to natural language processing (NLP), like search or language translation, is often problematic for low-resource languages. Synthetic data generation has been successfully used to generate parallel training data for training deep learning models for neural machine translation. Generating synthetic data for machine learning There are several standard approaches for generating synthetic data. These include the following:
Types of synthetic data Synthetic data can be classified into different types based on their usage and the data format. Generally, it falls into one of two categories:
Popular types of synthetic data, classified according to the data type, include the following:
Synthetic text finds its use in applications like language translation, content moderation, and product reviews. Synthetic images are used extensively for purposes like training self-driving cars, while synthetic audio and video data is used for applications including speech recognition, virtual assistants, and digital avatars. Synthetic time-series data are used in financial services to represent the temporal aspect of financial data, like stock price. Finally, synthetic tabular data is used in domains like e-commerce and fraud. Techniques for generating synthetic data Generating synthetic data can be very simple, such as adding noise to data samples, and can also be highly sophisticated, requiring the use of state-of-the-art models like generative adversarial networks. In this section, you’ll review two chief methods for generating synthetic data for machine learning and deep learning applications. Statistical methods In statistics, data samples can be assumed to be generated from a probability distribution with certain characteristic statistical features like mean, variance, skew, and so on. For instance, in the case of anomaly detection, one assumes that the nonanomalous samples belong to a certain statistical distribution while the anomalous or outlier samples do not correspond to this data distribution. Consider a hypothetical machine learning example of predicting the salaries of data scientists with certain years of experience at top tech companies. In the absence of real-world salary data, which is a topic considered taboo, synthetic salary data can be generated from a distribution defined by the few real-world salary public reports on platforms like Glassdoor, LinkedIn, or Quora. This can be used by recruiters and hiring teams to benchmark their own salary levels and adjust the salary offers to new hires. Deep learning-based methods As the complexity of the data increases, statistical-sampling-based methods are not a good choice for synthetic data generation. Neural networks, especially deep neural networks, are capable of making better approximations of complex, nonlinear data like faces or speech. A neural network essentially represents a transformation from a set of inputs to a complex output, and this transformation can be applied on synthetic inputs to generate synthetic outputs. Two popular neural network architectures for generating synthetic data are variational autoencoders and generative adversarial networks, which will be discussed in detail in the next sections. Variational autoencoders Variational autoencoders are generative models that belong to the autoencoder class of unsupervised models. They learn the underlying distribution of a data set and subsequently generate new data based on the learned representation. VAEs consist of two neural networks: an encoder that learns an efficient latent representation of the source data distribution and a decoder that aims to transform this latent representation back into the original space. The advantage of using VAEs is that the quality of the generated samples can be quantified objectively using the reconstruction error between the original distribution and the output of the decoder. VAEs can be trained efficiently through an objective function that minimizes the reconstruction error. VAEs represent a strong baseline approach for generating synthetic data. However, VAEs suffer from a few disadvantages. They are not able to learn efficient representations of heterogeneous data and are not straightforward to train and optimize. These problems can be overcome using generative adversarial networks. Generative adversarial networks GANs are a relatively new class of generative deep learning models. Like VAEs, GANs are based on simultaneously training two neural networks but via an adversarial process. A generative model, G, is used to learn the latent representation of the original data set and generate samples. The discriminator model, D, is a supervised model that learns to distinguish whether a random sample came from the original data set or is generated by G. The objective of the generator G is to maximize the probability of the discriminator D, making a classification error. This adversarial training process, similar to a zero-sum game, is continued until the discriminator can no longer distinguish between the original and synthetic data samples from the generator. GANs originally became popular for synthesizing images for a variety of computer-visionproblems, including image recognition, text-to-image and image-to-image translation, super resolution, and so on. Recently, GANs have proven to be highly versatile and useful for generating synthetic text as well as private or sensitive data like patient medical records. Synthetic data generation with Openlayer Openlayer is a machine learning debugging workspace that helps individual data scientists and enterprise organizations alike to track and version models, uncover errors, and generate synthetic data. It is primarily used to augment underrepresented portions or classes in the original training data set. Synthetic data is generated from existing data samples, and data-augmentation tests are conducted to verify whether the model’s predictions on the synthetic data are the same as for the original data. Conclusion In this article, you learned about synthetic data for machine learning and deep learning applications. In the absence of real-world data, as well as other pertinent issues like privacy concerns or the high costs of data acquisition and labeling, synthetic data presents a versatile and scalable solution. Synthetic data has found mainstream acceptance in a number of domains and for a variety of data types, including text, audio, video, time series, and tabular data. You explored these different types of synthetic data and the various methods for generation. These include statistical approaches as well as neural network–based methods like variational autoencoders and generative adversarial networks. Then you walked through a brief tutorial for generating synthetic data using deep learning methods. Finally, you saw the utility of third-party synthetic data generation products such as Openlayer, which can help companies rapidly scale their synthetic data requirements and accelerate model development and deployment. Related Blogs Published by Unbox.ai Data drift refers to the phenomenon where the distribution of live, real-world data differs or “drifts” from the distribution of data used to train a machine learning model. When data drift occurs, the performance of machine learning models in production degrades, resulting in inaccurate predictions. This reduction in the model’s predictive power can adversely impact the expected business value from the investment in training. If data drift is not identified in time, the machine learning model may become stale and eventually useless.
In this article, you’ll learn more about data drift, exploring why and in what ways it occurs, its impact, and how it can be mitigated and prevented. The Importance of Detecting Data Drift Machine learning models operate in a dynamic environment but are trained on data from a fixed, statistical distribution. Data drift can occur due to a variety of reasons, including seasonal variations, new product features, changes in customer behavior, or even rare events like the Covid-19 pandemic. Data drift is a critical challenge for production machine learning systems. It occurs when the statistical distribution of the target or real-world data diverges significantly from the statistical properties of the data on which the model was trained. This hurts model performance on new unseen data during real-world inference, leading to inaccurate predictions, poor customer experience, and monetary and reputational costs for the business. If undetected, data drift can cause multiple problems besides the obvious loss in model performance. It leads to greater MLOps challenges and technical burden for teams such as identifying data drift, conducting root cause analysis for input features correlated with data drift, data labeling, active learning, retraining, and redeploying the updated models to production. This is a significant investment of time and resources that can be avoided if machine learning models are closely monitored and a strategy for detecting and fixing data drift is in place. How to Identify Data Drift It is common to assume that a loss in model performance may be due to data drift. However, before arriving at this conclusion, it is important to assess data quality. Target data distributions could change due to a new set of users, a feature or product update, or even something as simple as a bug or formatting error in the code or data. After data-quality issues are ruled out, data drift can be examined in more detail. Fundamentally, data drift implies a change in the statistical distribution of target data from that of the training data. Thus, the simplest way to identify data drift is to compare summary statistics (like mean, variance, Kullback–Leibler divergence, and so on) of a carefully sampled subset of the target data against the training data. Other statistical measures include comparing the number of outliers in the two distributions or using the Kolmogorov–Smirnov test. Analyzing the correlations between input features and model predictions for both the data distributions can also shed some light. Model-based machine learning techniques can also be used to identify data drift. A sample of data from the reference or training distribution can be labeled as 0, and an equivalent sample from the target distribution can be labeled as 1. Based on this input data, a simple binary classification model can be trained to discriminate between the two types of data distributions. If the model can distinguish between the two data sets, this implies that data drift is present. Alternatively, if the model fails to discriminate between the data sets, then no data drift is evident. Using a machine learning–based approach captures nonlinear relationships better and can help catch data drift where the above statistical methods might fail. What are the different kinds of Data Drift? The change in the statistical distribution between the target and training data manifests in different forms of data drift that are observed in real-world machine learning systems: Covariate shift (feature drift) Covariate drift refers to a data drift that is correlated with a shift in the independent variables or input features. The relationship between the features and target variables is unchanged, but the change in a few features leads to covariate drift. Covariate drift can occur due to sample selection bias and is frequently observed in nonstationary environments. Concept drift Concept drift is associated with a change in the relationship between the independent variables and the target variables. For instance, a particular machine learning model may suffer from concept drift when it is launched in a new geography where the customer behavior is markedly different from the behavioral data from the original geography that was used to train the models. Although the set of input features and the data distributions may remain the same, the model may not make any useful predictions and is rendered obsolete. How can you mitigate and prevent data drift? Once data drift is confirmed as real and significant after rigorous analysis and statistical tests as described above, it is important to address it sooner than later. Here are a few strategies for doing so. Data labeling Labeling the new target data is the first step toward addressing data drift. A carefully selected batch of the test data can be sampled and sent to subject matter experts for data annotation. Thereafter, this labeled target data from the modified data distribution can be incorporated into the original data distribution to ameliorate the impact of data drift. Periodic model training With newly labeled target data, the model can be retrained on data from both the original distribution and the test distribution. As the new model is now trained to recognize data from the modified target distribution, it typically does a better job in production than the original model. However, a model might need to be trained multiple times, depending on the rate of data drift, to capture the new patterns from the test data. Model recalibration With repeated model retraining, the training pipeline, model architecture, and hyperparameters may remain the same, with the only difference being the change in training data. However, if data drift is not taken care of with periodic retraining, it might be prudent to train the model from scratch with a fresh approach and insights learned from efforts focused on evaluating and mitigating data drift. The new model may be trained differently from the original one in a number of ways:
Continuous monitoring Continuous monitoring of machine learning model performance is critical to keep track of the quality of the model in production. Model performance metrics like true positives, false positives, precision, recall, F1-score, and AUC-ROC curves can be periodically assessed. After thresholds for such performance metrics are carefully selected, alerts can be triggered using platforms like Grafana or Prometheus or by using third-party-managed MLOps platforms. Apart from output metrics, other things to monitor include any data issues or inconsistencies, bias in training data, and explainability metrics. Conclusion The phenomenon of data drift afflicts most machine learning models in production, arising from various reasons due to the dynamic nature of real-world data, seasonal trends, changes in product features, software- or data-related issues, changes in customer behavior due to new competition or legislation, and even rare black swan events like the Covid-19 pandemic. Data drift can be of different types, depending on whether the relationship between the independent features and the target variables changes or not. This article has equipped you to know what data drift looks like and provided a list of best practices for identifying and mitigating it before it becomes a major MLOps challenge and renders the machine learning model unfit for its intended business purpose. Related Blogs Published by Colabra Introduction
Effective communication skills are pivotal to success in science. From maximizing productivity at work through efficient teamwork and collaboration to preventing the spread of misinformation during global pandemics like Covid19, the importance of strong communication skills cannot be emphasized enough. However, scientists often struggle to communicate their work clearly for various reasons. Firstly, most academic institutes do not prioritize training scientists in essential soft skills like communication. With negligible organizational or departmental training and little to no feedback from professors and peers, scientists fail to fully appreciate the real-world importance and consequences of poor communication skills. The long scientific training period in the academic ivory tower is spent conversing with fellow scientists, with minimal interaction with non-technical professionals and the general public. Thus, the lingua franca among scientists is predominantly interspersed with jargon, leading to poor communication with non-scientists. This article will describe best practices and frameworks for professional scientists and non-scientists in commercial scientific enterprises to communicate effectively. How should scientists speak with non-scientists? IndustryThis section describes how professional scientists in industries like biotech and pharma can communicate better with cross-functional stakeholders from non-technical teams like sales, marketing, legal, business, product, finance, accounting, etc. Cross-functional collaboration In industry, scientists are often embedded in self-contained business or product teams with different roles. Taking a biotech product to market like a new drug, which has a long development cycle, involves extensive collaboration between specialists from multiple domains: research, quality assurance, legal and compliance, project management, risk and safety, vendor and supplier management, sales, marketing, logistics, and distribution, to name a few. Scientists are involved from the beginning of the process. However, scientists are often guilty of focusing solely on R&D without acutely considering how the science and technology underlying the product or business is operationalized by cross-functional teams and delivered to the market. Scientists are often less aware of the practical challenges of taking a drug prototype to the patient, such as long timelines due to multiple steps like risk management, safety reviews, regulatory approvals, coordination with pharmaceutical and logistics companies, and bureaucratic hurdles with governments and international bodies. This is a vital mistake in collaborative industry environments and often leads to poor job experience for scientists and their non-scientist peers and managers. The image below shows several communication challenges at the different stages of the drug development process that hinder successful commercialization. Although the various specialists share a common objective, each domain expert speaks a different “language” influenced by their respective training and fails to translate their opinions and concerns into a common language that all can understand. This comes in the way of optimal decision-making resulting in projects that stall even before demonstrating clinical efficacy. In an industry with a 90% drug development failure rate, poor communication and collaboration can be very expensive, to the tune of USD 1.3 billion per drug. The right culture is crucial to ensure successful outcomes, as advocated by AstraZeneca after a thorough review of their drug development pipeline. A recent real-world example pertains to the development of the AstraZeneca Covid-19 vaccine by multiple teams at the University of Oxford. Although the vaccine was developed within two weeks by February 2020, it was not until 30 December 2020 that the vaccine was finally approved for use in the UK, and it is even to date not authorized for use in the US. In particular, the AstraZeneca vaccine was subject to misinformation, fake news, and fear-mongering, which led to vaccine hesitancy and a lack of public trust. This led Drs. Sarah Gilbert and Catherine Green, co-developers of the vaccine, to author ‘Vaxxers,’ with the primary motivation to allay fears and reassure the general public about its safety and efficacy by explaining the science and process of creating the vaccine. Stakeholder management Another critical aspect of working with cross-functional teams involves managing key stakeholders to ensure a successful outcome for the project. Stakeholders often come from diverse non-scientific backgrounds, making working with them more challenging for scientists. The main challenge in effective stakeholder management is understanding the professional goals, metrics, and KPIs that drive each stakeholder. For instance, a product manager might focus on metrics like cost improvement over time, risk mitigation, or timelines; a finance leader may be focused on revenue; a compliance manager may be focused on metrics that capture safety and legal aspects. Understanding each cross-functional stakeholder’s north star can help scientists navigate the intricacies of stakeholder management. Effective stakeholder management involves numerous aspects: Identifying stakeholders The first step is to identify the stakeholders that are critical to the success of the scientific product and understand their motivations and priorities. Successful stakeholder management starts by mapping your stakeholders across several dimensions, including:
Aligning stakeholders Conflicting priorities among stakeholders are common and need to be resolved delicately. Achieving multi-stakeholder alignment for complex projects requires carefully planned discussions and negotiations to assess the lay of the land with each stakeholder and preempt potential conflicts. Focused group meetings that prioritize key points of disagreement or conflicting priorities can help achieve alignment and avoid conflicts. Engaging stakeholders After getting all the stakeholders aligned, it is useful to build a communication strategy to share project updates regularly. The communication plan must be tailored to each stakeholder. For example, individual contributors might need a high-touch approach, while project coordinators and administrators might just want periodic updates and high-level presentations. During the project's execution phase, continuous engagement and clear communication with the stakeholders are essential to keep everyone on the same page. Stakeholders may be involved in multiple biotech projects in parallel, and your project may not be their sole focus or priority. We have previously written about several modes of communication and project management apart from one-on-one meetings. At a minimum, it is beneficial to maintain a project status board detailing the progress of each milestone, metric, team, and timeline, especially to serve as a single source of truth, especially if some teams are working remotely. Entrepreneurship This section will discuss how aspiring startup founders with a scientific background should communicate and “sell” the company's mission to varied stakeholders from investors, employees, vendors, potential hires, and so on. Scientists with domain expertise and an entrepreneurial mindset are increasingly opting to build deep-tech startups soon after graduating from academia. From Genentech to Moderna and CRISPR Therapeutics to BioNTech, there is no shortage of successful biotech companies founded by scientists. However, building a commercially successful and viable biotech startup requires diverse skills with a much stronger need for excellent communication skills. Scientist founders need to have exceptional communication and sales skills to pitch the company to raise venture capital, write scientific grants, forge business partnerships with other companies, retain customers, attract talented employees with their vision for the company, give media interviews, and shape a mission-oriented organizational culture. Scientist-founders must communicate particularly well to bridge the gap between scientific research and commercialization. How should non-scientists speak with scientists? In this section, we will consider the viewpoint of non-scientists and how they can communicate more effectively with scientists. Non-scientists are typically more focused on product, business, sales, marketing, and related aspects of commercializing scientific research. The stakes for effective communication between scientists and managers are very high. This is best highlighted by NASA’s missions, which involve a diverse set of experts, both scientific and non-scientific, similar to the highly complex and multi-year projects described in the previous section. NASA’s failures on projects like the Columbia mission have been attributed to deficiencies in communication and insular company culture. Namely, management not heeding the scientists' and engineers’ warnings. These communication failures are expertly documented in a post-hoc report by the Columbia Accident Investigation Board – "Over time, a pattern of ineffective communication has resulted, leaving risks improperly defined, problems unreported, and concerns unexpressed," the report said. "The question is, why?" (source) Unfortunately, this state of affairs rings true even today in high-stakes and complex scientific enterprises. Here are some recommended tips that follow from such catastrophic mishaps and failures in workplace communication:
How can non-scientists better engage scientists? Non-scientist stakeholders' work largely focuses on business metrics, product roadmaps, customer research, project management, etc. These are critical focus areas that non-scientists need to update and communicate clearly to their scientist colleagues. In industry, it is common to observe scientist colleagues not actively participating in discussions focused on business topics and switch off until their work is the topic of discussion. It is crucial to engage scientists as they are on the front lines of core product development and in a better position to understand and flag potential roadblocks in manufacturing, commercialization, and logistics based on prior experience. Many product-related issues and bugs that surface later in the development cycle can be caught and addressed if there is more proactive communication between scientific and non-scientific teams. Scientists are generally trained to be conservative, focusing on accuracy and reliability, which can conflict with a manager’s ambitious goals for time-to-market or revenue targets. In these situations, managers should allow scientists to voice their concerns, not be afraid to dive deeper, coordinate with other cross-functional stakeholders, and take a balanced decision integrating every stakeholder’s views. In the long term, cultivating an open and progressive culture that encourages debates and tough discussions reaps enormous benefits whereby no business-critical concern is left unvoiced. A transparent and meritocratic culture promotes greater cooperation and understanding among different teams striving towards the same goals. Conclusion We discussed why scientists often struggle with effective communication with other scientists and non-scientist stakeholders when working in industry or building their own company. We addressed how scientists should approach communication with non-scientist colleagues and how to collaborate with them. We also discussed effective communication strategies from the perspective of non-scientists speaking to scientists. In the long run, having strong communication and soft skills confers greater career durability than simply having scientific and technical skills. Understanding this and upskilling accordingly can empower scientists to transition and perform well in industry. Related Blogs Published by Unbox.ai Introduction
Supervised machine learning models are trained using data and their associated labels. For example, to discriminate between a cat and a dog present in an image, the model is fed images of cats or dogs and a corresponding label of “cat” or “dog” for each image. Assigning a category to each data sample is referred to as data labeling. Data labeling is essential to imparting machines with knowledge of the world that is relevant for the particular machine learning use case. Without labels, models do not have any explicit understanding of the information in a given data set. A popular example that demonstrates the value of data labeling is the ImageNet data set. More than a million images were labeled with hundreds of object categories to create this pioneering data set that heralded the deep-learning era. In this article, you’ll learn more about data labeling and its use cases, processes, and best practices. Why is data labeling important? Labeled data is necessary to build discriminative machine learning models that classify a data sample into one or more categories. Once a machine learning model is trained using data and corresponding labels, it can predict the label of a new unseen data sample. Data labeling is a crucial process as it directly impacts the accuracy of the model. If a significant proportion of the training data set is mislabeled, it will cause the model to make inaccurate predictions. Data labeling of production data is also important to counter data drift. The model can be continuously improved by incorporating the newly labeled samples from the real-world data distribution into the training data set. Poorly labeled data can also introduce bias in the data set, which can cause the models to consistently make inaccurate predictions on a subset of real-world data. Mislabelingcan severely impact the fairness and accuracy of models and warrants additional efforts to detect and eliminate labeling errors. Relabeling helps to address mislabeled samples, improving the data quality and, consequently, the accuracy of the machine learning models. How is data labeling performed? Again, data labeling helps train supervised machine learning models that learn from data and their corresponding labels. For example, the following text, sourced from the Large Movie Review Dataset, can be annotated in a number of ways depending on the use case: I saw this movie in NEW York city. I was waiting for a bus the next morning, so it was 2 or 3 in the morning. It was raining, and did not want to wait at the PORT AUTHORTY. So I went across the street and saw the worst film of my life. It was so bad, that I chose to stay and see the whole movie,I have yet to see anything else that bad since. The year was 69,so call me crazy. I stayed only because I could not belive it.........1. Use case: Sentiment analysis
For the named entity recognition use case, data annotators have to review the entire text and identify and label any mention of places. Typically, data annotation is outsourced to vendors who contract subject matter experts relevant for the specific machine learning use case. The team of annotators are assigned different batches of data to label on a daily basis for the duration of the project, using simple tools like Excel or more sophisticated labeling platforms like Label Studio. Labelers’ performance is evaluated in terms of metrics like overall accuracy and throughput—i.e., the number of samples labeled in a day. If the same set of data samples are assigned to multiple annotators, then the labels given by each annotator can be combined through a majority vote. Inter-annotator agreementhelps to reduce bias and mislabeling errors. For several use cases, data labeling can be extremely painstaking and time-consuming, which may lead to labeling fatigue. To counter this, labels assigned to each annotator undergo one or more rounds of review to catch any systematic errors. Once a batch of data is labeled, reviewed, and validated, it is shared with the data science team, who review select samples for labeling accuracy and verification and then provide feedback to the annotators. This iterative and collaborative process ensures that the final labels are of high quality and accuracy to use for training machine learning models. How is data relabeling performed? The repetitive and manual nature of data labeling is often fraught with errors. This necessitates the need to identify and relabel samples that were erroneously labeled the first time around. Relabeling is an expensive but necessary process as it is imperative to have a training data set of high quality. Unlike labeling, relabeling is usually done on a smaller sample of the entire data set and can be completed much faster if the samples are mislabeled in a unique way or associated with the same annotator. Once a trained model is deployed, its predictions on real-world data can be evaluated. A detailed error-analysis process can sometimes reveal systematic prediction errors. Many times, these characteristic errors may be correlated with a certain type of data sample or feature. In such cases, having another look at similar samples in the training data can help identify mislabeled samples. More often than not, labeling errors on a certain segment of the training data can be captured through such error analysis and corrected with relabeling. Best practices for data labeling Data labeling can be prohibitively expensive and time-consuming for large data sets. As model development is contingent on the availability of good-quality labeled data, poor labeling can affect the timelines and prolong the time to build and deploy machine learning models. A good practice for data scientists is to curate a comprehensive data-annotation framework for each use case before starting the data-labeling process. Clear, structured guidelines with examples and edge cases provide much-needed clarity for annotators to do their job with greater speed and accuracy. In the absence of domain experts within the company, external experts can be sought to discuss and conceptualize guidelines and best practices for labeling specific types of data. As labeling of large data sets by domain experts can be quite expensive, in specific cases, data labeling can be crowdsourced to thousands of users on platforms like Amazon Mechanical Turk. Typically, labeling by crowdsourced users is fast but often noisy and less accurate. Still, crowdsourcing can be a significantly quicker method of collecting the first set of labels before doing one or more rounds of relabeling to eliminate errors. Error analysis is another recommended practice to diagnose model prediction errors and iteratively improve model performance. Error analysis can be done manually by the data scientists or with greater speed and reproducibility using machine learning debugging platforms like Openlayer. Another good practice, in the context of very large data sets for deep learning applications, is to leverage machine learning to obtain a first pass of labels using techniques like the following: Conclusion Machine learning and deep-learning models are typically trained on large data sets. To train such models, a label for each data sample is necessary to teach the model about the information in the data set. Labeling, therefore, is an integral aspect of the machine learning lifecycle and directly influences the quality and performance of models in production. In this article, you’ve seen the importance, process, and best practices for efficient data labeling and relabeling. Mislabeled data samples introduce noise and bias in the data set that adversely impact the performance of the model. Identifying mislabeled examples through error analysis is a proven technique to improve the quality of training data that can be accelerated using machine learning debugging and testing platforms like Openlayer. Related Blogs Published by Unbox.ai Introduction
Modern companies now unanimously recognize the value of data for driving business growth. However, high-quality data is much more valuable than data assets of poor quality. As companies accumulate petabytes of data from various sources, it becomes imperative to focus on the quality of data and filter out bad data. Data is the fundamental building block for predictive machine learning models. Although having access to greater amounts of data is beneficial, it doesn’t always translate to better-performing machine learning models. Sampling training data that passes quality checks and meets certain acceptance criteria can significantly boost the accuracy of the model predictions. In this article, you’ll learn more about why high-quality data is essential for building robust machine learning models, expanding on the various parameters that define data quality: accuracy, completeness, consistency, timeliness, uniqueness, and validity. You’ll also explore a few mechanisms you can implement to measure and improve the quality of your data. What is data quality? Data quality is a measure of how suitable the data is for its intended applications in data analytics, data science, or machine learning. There are several dimensions along which data quality is measured, which include the following:
Why is data quality important? Data quality is an important determinant of the quality of decision-making within an organization. Poor-quality data leads to inaccurate analytics and machine learning models, which might adversely impact various business operations as well as customer experience. Decisions and business strategies based on flawed data can have massive consequences. Typical data-quality issues include data security and data that is incomplete, duplicated, inconsistent, incorrect, missing, poorly defined, poorly organized, or stale. In the context of data science use cases, the consequences of using poor-quality data can be immense—machine learning models trained on low-quality data invariably generate weak or inaccurate predictions, which are not easy to troubleshoot. Deep-learning models in particular are very data-hungry, and their state-of-the-art performance is driven by the massive amounts of data on which they are trained. In this context, recent work has shown that training models with less data reflects real-world scenarios better and is increasingly becoming the norm. The cost of bad data to organizations is also enormous—as per an IBM study, the yearly cost of poor-quality data in the US alone is equal to USD 3.1 trillion. Therefore, it is paramount for organizations to invest in proper measurement and evaluation of data quality before building data-driven applications or devising new business strategies. Determining data quality Several organizations, from IMF to World Bank, have formulated Data Quality Assessment Frameworks (DQAF) to establish clear guidelines for measuring the quality of data in terms of accuracy, completeness, consistency, timeliness, uniqueness, and validity. This section will focus on each of these data-quality dimensions and discuss how they define the quality of data. Accuracy Accuracy, as the term implies, is a pivotal aspect of data quality—it means that the information is correct. Naturally, inaccurate information can cause many significant problems for a business. For instance, consider an example in which the time of financial transactions is incorrectly recorded due to a failure to update to daylight saving time. In such a scenario, the timing offset could lead to inaccurate analysis and reporting of core business metrics like daily sales and revenue. Such data inaccuracies can lead to potentially damaging consequences of incorrect financial and tax filings that could result in financial penalties by regulatory bodies. Completeness Completeness refers to how comprehensive the data is and whether it contains all the fields and values necessary to make them fit for the intended purpose. Incomplete data often contains empty or missing values across rows or columns and is unusable for further analysis. For instance, if a customer’s email address is missing, then this customer may not feature in any marketing campaigns, resulting in a potential loss of business for the company. Consistency Consistency is another fundamental trait of data quality, as it can affect the usage of the entire data set. If a data set has millions of records but some rows store a customer’s name as “CustomerName” while the remaining rows store the same information as “FirstName” and “LastName” separately, it might lead to inaccurate results and analysis. Another common example of inconsistent data is related to the underlying format or units of specific data fields. For instance, data like time is often kept in inconsistent formats, and units of money may be recorded differently from country to country. Timeliness Timeliness refers to how recent and up-to-date the information is. For a number of applications, timely data is essential as it captures the current trends and patterns in customer behavior or business health. Data tends to lose its value over time and can drastically affect the quality of business decisions as well as predictions from machine learning models trained on older data. It can cost organizations lost time and money, in addition to reputational damage. Uniqueness Uniqueness refers to the lack of duplication or overlap within a data set or across data sets. Modeling redundant information can often lead to spurious correlations or results that can adversely affect statistical analysis as well as model predictions. Thus, uniqueness is a critical dimension of data quality that is important to build trust in the data for downstream use cases. Validity For several data fields, validation checks are important. For instance, a mobile phone number is usually ten digits long, and zip codes in the US should have five digits. When data does not conform to standard formats or business-specific rules, it is said to be invalid. Invalid data can cause grave errors in downstream analytics and necessitates careful scrutiny of every data column before using it. Truncation of data also leads to data-validity problems. For instance, a user may mistakenly input six digits for a US zip code, which gets truncated to five digits. While such an input may pass data-validation checks, it is ultimately inaccurate. Additional sources of data-validity errors arise due to mismatched data formats. For instance, a data type like zip code may be inconsistently saved in numeric or string format. Improving data quality There are numerous methods for improving data quality. The first step often involves data profiling—that is, doing an initial assessment of the current state of the data sets. Defining what is good data is also critical to establishing guardrails around selecting data for further usage. Furthermore, a number of checks for data validation, completeness, consistency, and timeliness can be defined and have to be met by all current and new data sets. Data standardization across the organization helps to meet data-quality standards so that every stakeholder across different divisions has the same understanding of the various data sets and fields. Implementing a robust data governance framework can also help businesses improve the quality of organizational data. Finally, recent advances in machine learning and deep learning can also be used to identify and improve the quality of data in a more scalable and reproducible fashion. For example, in the deep-learning study, a data-quality assessment framework grounded in statistics and deep learning was used to identify outliers in a data set of salary information published by the state of Arkansas, USA. As the size of organizational data is bound to increase exponentially in the coming years, companies ought to allocate dedicated resources and investments in new techniques from fields like machine learning and deep learning to measure and provide statistical insights into the quality of their data. Conclusion In this article, you’ve learned what data quality is and why it is important for organizations to measure and evaluate the quality of their in-house data. Poor-quality data can have significant consequences for a business in terms of inaccurate analytics, predictive machine learning models trained on bad data, as well as ill-informed business decisions and strategies. Data quality can be measured in terms of a number of parameters such as accuracy, completeness, consistency, timeliness, uniqueness, and validity. Each of these data-quality dimensions are important, and organizations can improve the quality of their data by having robust data profiling, standardization, and validation checks in place. More recently, advances from machine learning and deep learning can also be harnessed to quantitatively define and evaluate the quality of data. Related Blogs
Published by Transform Introduction
A metric layer is a centralized repository for key business metric. This “layer” sits between an organization’s data storage and compute layer and downstream tools where metric logic lives—like downstream business intelligence tools. A metric layer is a semantic layer where data teams can centrally define and store business metrics (or key performance indicators) in code. It then becomes a source of truth for metric—which means people who analyze data in downstream tools like Hex, Mode, or Tableau will all be working with the same metric logic in their analyses. The metric layer is a relatively new concept in the modern data stack, mainly because until recently, it was only available to companies with large or sophisticated data teams. Now it is more readily available to all organizations with metric platforms like Transform. In this article, you’ll learn what a metric layer is, how to use your data warehouse as a data source for the metric layer, and how to get value from this central metric repository by consuming metrics in downstream tools. How a Metric Layer fits into a Modern Data StackThe modern data stack is composed of a number of elements organized in the order of how data flows:
One central benefit of a metric layer is that it sits between the data warehouse and downstream analytics tools. People can access metrics in business intelligence (BI) tools like Tableau, Mode, and Hex, bringing metrics consistency across all business analysis. Use cases for the Metric Layer The formulation and implementation of metric layers was pioneered by prominent tech companies like Airbnb, Spotify, Slack, and Uber. Airbnb designed a metric layer called Minerva to serve as a single source of truth (SSOT) metric platform. They did this by standardizing the way metrics are created, calculated, served, and used across the organization. Uber built uMetric, a standardized metric platform that underlies the entire lifecycle of a metric from definition, discovery, planning, calculation, quality, and consumption. These pillars not only enable rapid metric computation for business decisions, but also help create useful features for training ML models and promoting data democratization. A new component in the Modern Data StackWith the emergence of big data, predictive analytics, and data science, most companies have access to enormous amounts of valuable data. Many organizations have evolved their data stack to simplify computation, transformation, and access to key business metrics, which can accelerate data-driven decision-making. However, as Benn Stancil noted in his popular Substack blog, there was no central repository for defining metrics. This causes confusion and misalignment across an organization. "The core problem is that there’s no central repository for defining a metric. Without that, metric formulas are scattered across tools, buried in hidden dashboards, and recreated, rewritten, and reused with no oversight or guidance." —Benn Stancil, The missing piece of the modern data stack Another common issue is “dashboard sprawl” where metric logic is spread across different tools and data artifacts. Since this logic is different for every tool, teams often end up with different numbers for the same metrics and no one knows where to find the “correct” metric to answer their most important business questions. This problem led to the metric layer becoming a new artifact in the modern data stack. With a single shared store of metrics definitions and values, the metric layer ensures consistent and accurate analysis and reporting of metrics. A metric layer not only centralizes key business data but also helps improve the efficiency of data teams by removing the need for repeated analytics. This helps data stakeholders become key advocates and enablers of data-driven decision-making and data democratization across the entire organization. Reutilization of metrics in diverse contexts and external tools One of the benefits of having a single metrics repository is that it can be connected to a variety of tools; for example, CRM’s, BI tools, tools developed in-house, as well as data quality and experimentation tools. A centralized architecture ensures that no matter how a tool’s internal logic is configured, the end result will be based on the same metric logic and consistent across tools and applications. For instance, MetricFlow, the metric layer behind Transform, has an API that enables users to express requests for their Transform metrics directly within SQL expressions. Core metrics like Net Promoter Score (NPS), Monthly Recurring Revenue (MRR), Customer Acquisition Cost (CAC), loan-to-value (LTV), and Annual Recurring Revenue (ARR) capture the health of the business and need to be accurate for reporting and decision-making. With a metric layer, it’s possible to see the lineage of each metric, how it’s built, what the data source is, and how it’s consumed. By unifying metrics extraction and data analytics on these metrics, the metric layer provides the much-needed consistency that is lacking in modern data stacks. Enhancing transparency between technical and non-technical teams with a single interface A single interface for metrics information gives data stakeholders across an organization—in development, sales, marketing, and more—to have the same view and understanding of key metrics to track goals. This consistency allows all of these teams to speak the same language regardless of the tools they use to compute the metrics. This is a tremendous benefit of a metric layer and promotes stronger data democratization and governance across the entire organization. Transform is unique in that it has the addition of a metrics catalog on top of MetricFlow, its open source metric layer. The metrics catalog is a central location where both data teams and non-technical users can interact with, build context, collaborate on, and share key metrics. Tracking changes is easier Because businesses are constantly evolving and creating new metrics or changing the definition of existing metrics, each data stakeholder has to manually keep track of changes in a data warehouse to update their metrics definition and logic. However, with the combination of a metric layer and a metrics catalog, tracking changes metrics owners are alerted anytime the lineage or definition of a metric changes. This enables data stakeholders to make better sense of data, especially when a new metric definition leads to anomalous or unexpected results. Dig into the Metric Layer A metric layer reduces the problem of disparate results when the same metric is computed by different teams using a wide variety of BI tools. And it makes data-driven analytics more precise and promotes faster and more accurate decision-making. If you’re looking for a streamlined and centralized metric layer, MetricFlow is now open source. You can explore the project on Github. Find more information about Transform’s metric layer and its benefits in the product documentation. Related Blogs
I receive several messages about the benefits of joining FAANG and similar companies and startups in the context of Data Science, Machine Learning & AI roles.
Here’s my take, in no particular order: 1. 𝐁𝐫𝐚𝐧𝐝. FAANG+ are not only the top technology companies but also the biggest companies by market cap -> great brand to add to your profile, top compensation and benefits. 2. 𝐒𝐜𝐨𝐩𝐞. The scope of AI/ML applications in these companies is tremendous as they have tons of data. You can get to work on multiple use cases, driven by statistics, machine learning, deep learning, unsupervised / semi-supervised / self-supervised, reinforcement learning etc. Internal team transfers facilitate expanding your breadth of ML experience. 3. 𝐁𝐚𝐫. The AI/ML work is cutting edge, as most of these companies invest heavily in R&D and create game-changing techniques and models. They also invest heavily in platform, cloud, services etc. that make it easier to build and deploy ML products. 4. 𝐑&𝐃. You can do both research on moon-shot projects if that’s your cup of tea, as well as more immediate business-driven data science projects with monthly or quarterly deliverables. 5. 𝐏𝐞𝐨𝐩𝐥𝐞. You get to work with the creme-de-al-creme in terms of talent, ideas, vision, and execution. Your own level will rise if you are surrounded by some of the brightest folks, and also get to collaborate with their clients and collaborators from academia, startups as well. 6. 𝐍𝐞𝐭𝐰𝐨𝐫𝐤. After FAANG, people go on to do many diverse things — from building a startup to doing cutting-edge research to non-profits to venture capital amongst others. You can find quality partners for the next steps of your career journey. 7. 𝐒𝐲𝐬𝐭𝐞𝐦𝐬. Processes and systems for AI/ML/Data are more mature and streamlined than smaller/newer companies which can facilitate your speed and execution of your projects. 8. 𝐂𝐮𝐥𝐭𝐮𝐫𝐞. The culture, on average, is more professional as these companies invest heavily in their employees and regularly come up with new employee-friendly policies to make it a great place to work. 9. 𝐅𝐫𝐞𝐞𝐝𝐨𝐦. After FAANG, you will be in demand and recruiters and hiring managers will seek you out if you’ve proved your chops whilst at the company. You will have more opportunities to sample from and greater freedom in terms of deciding your career and life trajectory, as you can also move internally to different countries. 10. 𝐈𝐦𝐩𝐚𝐜𝐭. Given the scale at which these companies operate, the scope for real-world measurable impact is enormous. There are some downsides, caveats and exceptions as well, but on average these factors make FAANG and similar tech companies a very attractive proposition to launch, build and grow your career in data science and machine learning. Introduction
"Data democratization" has become a buzzword for a reason. Modern organizations rely extensively on data to make informed decisions about their customers, products, strategy, and to assess the health of the business. But even with an abundance of data, if your business can’t access or leverage this data to make decisions, it’s not useful. To that end, data democratization, or the process of making data accessible to everyone, is quintessential to data-driven organizations. Providing data access to everyone also implies that there are few if any roadblocks or gatekeepers who control this access. When stakeholders from different departments—like sales, marketing, operations, and finance—are permitted and incentivized to use this data to better understand and improve their business function, the whole organization benefits. Successful data democratization requires constant effort and discipline. It’s founded on an organization-wide cultural shift that embraces a data-first approach and empowers every stakeholder to comfortably use data and make better data-driven decisions. As Transform co-founder James Mayfield put it, organizations should think about "democratizing insights, not data." In this article, I will provide a detailed overview of data democratization, why organizations should invest in it, and how to actually implement it in practice. Why democratize access to data? Historically, data used to be kept in silos, usually under the purview of the IT or Analytics departments. When any stakeholder from outside these departments required data for their work, they had to go through these data gatekeepers to access the necessary assets. This philosophy has been the norm for decades but is no longer relevant for modern data-driven organizations. Removing these types of bottlenecks is a necessary first step toward data democratization. Guidelines for data democratization can be noted in a data governance framework to improve access and provide high-quality data for downstream analytics. Improving access is just the first step of an ongoing process where every individual employee is encouraged and trained to make use of data. The more people who can make decisions based on data, the more the organization stands to benefit from a variety of perspectives and ideas. Companies have been dedicating huge investments in data infrastructure and tooling in order to build an analytics advantage over their competitors. The dream is to “democratize data” and get employees to change their ways of working and start making decisions informed by data, not gut feelings. By investing in data education and helping analysts influence, then building modern tools to support metrics, we will continue making progress toward that goal of truly democratized data" —James Mayfield, co-founder, Transform While data analytics and business intelligence efforts are traditionally the domain of data experts, organizations can empower non-technical stakeholders to perform basic data operations via in-house training programs, workshops, and self-service tools that can simplify their onboarding and learning process. They can also use software that surfaces data in an easy-to-consume format for business stakeholders. Data democratization has multiple downstream benefits. It leads to greater data literacy, which can facilitate not only greater data-driven decision-making but also potentially lead to creation of new products or services based on insights mined from the data. Therefore, greater democratization, usage, and adoption of a data-driven approach can unlock massive commercial value and new growth levers for businesses. How do you actually democratize data? Implementing data democratization is a hard challenge and an ongoing process. To be successful, it needs support, buy-in, and a lot of patience from the leadership. Apart from conceptualizing and implementing curated data governance frameworks and policies, organizations can leverage tools to enable data democratization at scale. Tools to enable data democratization The Data Catalog A data catalog is a collection of metadata that, combined with data management and search tools, helps data stakeholders find and acquire data for downstream analytics. A data catalog provides a managed and scalable data discovery and metadata management capabilities which are fundamental requirements of attaining higher levels of data democratization in an organization. The Data Mart A data mart is a subset of a data warehouse focused on a specific business vertical or data domain. Data marts enable specific users to access specific data that empowers them to quickly access these datasets without wasting time searching for the same in the data warehouse. For instance, individual departments like sales, marketing, operations, and finance can have their respective data marts for accelerating their domain-specific data-driven decision making. The Metrics Catalog A metrics catalog is a new layer in the modern data stack. It is a centralized store for all of your organizations’ most important metrics (or key performance indicators) and it's uniquely positioned between the data warehouse and downstream tools. As a self-service place for business KPIs, every stakeholder in the organization has access to track their own metrics and share context with others. By capturing core business metrics in this fashion and this location in the modern data stack, a metrics catalog provides immense visibility and transparency into an organization's most critical metrics and metric lineage for all stakeholders in an organization. This new concept of a metrics catalog can have a significant role to play in democratizing data to everyone. As a single source of ground truth for business data, a metrics catalog enables diverse stakeholders to base all key decisions on the same foundation. It also allows for disparate teams to use the same metrics, ask questions, and keep everyone aligned and on track. This greatly enhances the level of data democratization within an organization. Challenges for data democratization Although the benefits of data democratization are pretty evident, there are also numerous challenges. Some challenges are common, like data being kept in silos and unclear data ownership. The informational silos problem is antithetical to data democratization, and can adversely impact an organization's ability to leverage data for improvising its business performance and decision making. Different teams have ownership of different types of data, which contributes to the problem of information silos. When a particular team has exclusive access to specific data assets, they not only hinder other teams from accessing the data but also guard their analysis and insights derived from the same data. This often leads to duplication of efforts across teams, causing a massive waste of organizational time and resources. As each individual team or department hoards its own data and analyses, it contributes to the adoption of the same undemocratic processes across other teams further compounding the challenges in promoting data democratization. With greater access to the organizational data assets, there is also a challenge of data security, privacy, and potential misuse of the data. It increases the number of gaps in the organization which might become vulnerable to adversarial attacks and data breachers. This is why it’s important to have a balance between data security and data access—including having stronger safeguards around who can access and analyze personally-identifiable information and customer data. Looking ahead If implemented well, data democratization can provide an immense competitive edge that will only compound over time as organizations mature in their digital transformation journey. Several tools and data artifacts can aid in better implementation and adoption of best practices and policies that help in democratizing data. A metrics catalog is one relatively new tool that provides a centralized store of business critical information accessible to multiple stakeholders. It captures essential business metrics and provides a simplified interface that is agnostic of the separate analytics, CRM, and BI platforms used by various teams in the organization. Learn more about how a metrics store can promote data democratization and governance at Transform.co. Published by Neptune.ai Introduction
Machine learning and deep learning models are everywhere around us in modern organizations. The number of AI use cases has been increasing exponentially with the rapid development of new algorithms, cheaper compute, and greater availability of data. Every industry has appropriate machine learning and deep learning applications, from banking to healthcare to education to manufacturing, construction, and beyond. One of the biggest challenges in all of these ML and DL projects in different industries is model improvement. So, in this article, we’re going to explore ways to improve machine learning models built on structured data (time-series, categorical data, tabular data) and deep learning models built on unstructured data (text, images, audio, video, or multi-modal). The strategy for improving machine learning models At this point, implementing ML and DL applications in business is still in its early days, and there is no single structured process that can guarantee success. However, there are some best practices that can minimize the likelihood of a failed AI project [1, 2, 3]. One of the main keys to success is model accuracy and performance. Model performance is mainly a technical factor, and for a number of machine learning and deep learning use cases, deployment doesn’t make sense if the model isn’t accurate enough for the given business use case. In the context of improving existing machine learning and deep learning models, there’s no one-size-fits-all strategy that can be consistently applied. I will review a set of guidelines and best practices that can be evaluated to systematically identify potential sources of improvement in accuracy and model performance. Table 1, above, shows a set of high-level factors that should be considered before starting to debug and improve ML and DL models. It highlights the crucial set of factors that underlie the business and technical constraints within which the machine learning or deep learning model has to be improved. For example, a machine learning model for predicting credit rating of new retail banking customers should also be able to explain its decision in case the credit card application is rejected. Here, simply optimizing for the technical metric isn’t enough if the model doesn’t offer explainability and guidance for the customer to understand and improve their credit score. For clarity’s sake, in this article, I assume that your machine learning or deep learning model has already been trained on in-house data for a specific business use case, and the challenge is to improve the model performance on the same test set to meet the required acceptance criteria. We’re going to explore several methods to improve model performance, so you’ll surely find one or two relevant to your use case. Ultimately, practice and experience working on a wide variety of models leads to better intuition about the best approaches to improve model accuracy and prioritize these techniques over others. Preliminary analysis The first step in improving machine learning models is to carefully review the underlying hypotheses for the model in the context of the business use case, and evaluate the performance of the current models. (1) Review initial hypotheses about the dataset and the choice of algorithms In an ideal scenario, any machine learning modeling or algorithmic work is preceded by careful analysis of the problem at hand including a precise definition of the use case and the business and technical metrics to optimize [1]. It’s far too common to lose sight of the pre-defined data annotation guidelines, dataset creation strategies, metrics and success criteria once the exciting stage of building machine learning or deep learning models begins. However, keeping the larger picture in mind is beneficial to streamline and prioritize the iterative process of improving machine learning and deep learning models. (2) Is the model overfitting or underfitting? This can be visualized as in Figure 1, below, by plotting the model prediction error as a function of model complexity or number of epochs. The difference between the training and test error curves shows overfitting, i.e., high variance and low bias or underfitting, i.e., high bias and low variance, and provides a useful proxy to understand the current state of the machine learning model. If the model is overfitting, it can be improved by :
(3) What kind of errors is the model making? For a typical classification problem, this can be visualized using plots like the Confusion Matrix, which illustrates the proportion of Type 1 (false positive), and Type 2 (false negative) errors. Figure 2 shows a confusion matrix for a representative binary classification problem. In this example, we have 15 True Positives, 12 False Positives, 118 True Negatives, 47 False Negatives. So:
Here, both precision and recall are on the lower side, and depending on the acceptance criteria, there is ample scope to improve both. Depending on the business use case and domain, it might make more sense to focus on improving recall compared to precision. This applies to several machine learning problems in domains like healthcare, finance, and education. Figure 3 shows another representative confusion matrix for a multi-class classification problem, a common use case in industry applications. At first glance, it’s clear to see that the model is confusing classes 1-5 with class 0, and in certain cases, it’s predicting class 0 more often than the true class. This suggests that there’s a systematic error in the model, most likely to do with class 0. Armed with this insight, the first step to improve this model would be to check the labeled training examples for potential annotation errors or for the degree of similarity between the examples belonging to class 0 vs. classes 1-5. Potentially, this error analysis might show relevant evidence like the labels from a particular annotator being systematically mislabeled, accounting for the high confusion rate between the corresponding classes or categories. Model optimization After initial analysis and evaluation of model accuracy, visualization of key metrics to diagnose the errors, you should see if you can extract additional performance from the current model by retraining it with a different set of hyperparameters. The assumption underlying a trained machine learning or deep learning model is that the current set of model weights and biases correspond to a local minima during the convex optimization process. Gradient descent should ideally yield a global minima that corresponds to the most optimal set of model weights. However, gradient descent is a stochastic process that varies as a function of several parameters including how the weights are initialized, the learning rate schedule, the number of training epochs, any regularization method used to prevent overfitting, and a range of other hyperparameters specific to the training process and the model itself. Each machine learning and deep learning model is based on a unique algorithm and intrinsic parameters. The goal of machine learning is to learn the best set of weights to approximate complex nonlinear functions from data. It’s often the case that the first trained model is suboptimal and finding the optimal combination of hyperparameters can yield additional accuracy. Hyperparameter tuning involves training separate versions of the models, each trained on a different combination of hyperparameters. Typically, for smaller machine learning models, it’s a quick process and helps identify the model with the highest accuracy. For more complex models including deep neural networks, running several iterations of the same model on different combinations of hyperparameter values may not be feasible. In such cases, it’s prudent to limit the range and choice of individual hyperparameter values based on prior knowledge or existing literature to find the most optimal model. Three methods of hyperparameter tuning are most commonly used: (1) Grid Search Grid search is a common hyperparameter optimization method that involves finding an optimal set of hyperparameters by evaluating all their possible combinations. It’s most useful when the optimal range of relevant hyperparameters are known in advance, either based on empirical experiments, previous work, or published literature. For instance, if you have identified 6 key hyperparameters and 5 possible values for each hyperparameter within a specific range, then grid search will evaluate 5 * 6 = 30 different models for each unique combination of hyperparameters. This ensures that our prior knowledge about the hyperparameter range is captured into a finite set of model evaluations. The downside of this method is it’s computationally expensive and it only samples from well-defined spaces in the high-dimensional hyperparameter grid. Therefore, as shown in Figure 4, it’s more likely to miss the local minima associated with optimal hyperparameter values outside the pre-defined range. To alleviate these limitations of grid search, random search is recommended. (2) Random Search Random search essentially involves taking random samples of the hyperparameter values, and is better at identifying optimal hyperparameter values that one may not have a strong hypothesis about [4]. The random sampling process is more efficient and usually returns a set of optimal values based on fewer model iterations. Therefore, random search is the first choice for hyperparameter optimization in many cases. (3) Bayesian Search Bayesian search is a sophisticated hyperparameter optimization method based on the Bayes Theorem [5]. It works by building a probabilistic model of the objective function, called the surrogate function, that is then searched efficiently with an acquisition function before candidate samples are chosen for evaluation on the real objective function. Bayesian Optimization is often able to yield more optimal solutions than random search as shown in Figure 5, and is used in applied machine learning to tune the hyperparameters of a given well-performing model on a validation dataset. Models & algorithms (1) Establish a strong baseline model To improve your machine learning or deep learning model, it’s important to establish a strong baseline model. A good baseline model incorporates all the business and technical requirements, tests the data engineering and model deployment pipelines, and serves as a benchmark for subsequent model development. The choice of the baseline model is influenced by the particular application, the kind of dataset, and the business domain. For instance, for a forecasting application for time-series data from the financial domain, an XGBoost model is a strong baseline model. In fact, for several regression and classification based applications, Gradient Boosted Decision Trees are commonly used in production. Therefore, it makes sense to start with a model that is known to produce robust performance in production settings. For unstructured data like images, text, audio, video, deep learning models are commonly employed across applications like object classification, image segmentation, sentiment analysis, chatbots, speech recognition, emotion recognition amongst others. Given the rapid advancement in the state-of-the-art performance of deep learning models, it’s prudent to use a more sophisticated model compared to an older one. For instance, for object classification, deep convolutional network models like VGG-16 or ResNet-50 should be the baseline, instead of a single layer convolutional neural network. As an example, for a face recognition application for CCTV image data from the security domain, a ResNet-50 is a strong baseline contender. (2) Use pre-trained models and cloud APIs Instead of training a baseline model yourself, in certain cases, you can save valuable time and energy by evaluating pre-trained models. There are a variety of sources like Github, Kaggle, or APIs from cloud companies like AWS, Google Cloud, Microsoft Azure, specialized startups like Scale AI, Hugging Face, Primer.ai amongst others. The advantage of using pretrained models or APIs is ease of use, faster evaluation, and savings in time and resources. However, an important caveat is that such pretrained models are often not directly applicable for your use cases, less flexible, and tricky to customize. Using Transfer Learning, however, pretrained models can be applied to your use case by not retraining complex models afresh, and instead fine-tuning model weights on your specific dataset. For example, the intrinsic knowledge of an object classification model like ResNet-50 trained on several image categories from the ImageNet dataset can be leveraged to accelerate model development for your custom dataset and use case. APIs are available for numerous use cases like forecasting, fraud, search, optical character recognition for processing documents, personalization, chat and voice bots for customer service, and others [6]. (3) Try AutoML While pretrained models are readily available, you can also investigate state-of-the-art AutoML technology for creating custom machine learning and deep learning models. AutoML is a good solution for companies that have limited organizational knowledge and resources to deploy machine learning at scale to meet their business needs. AutoML solutions are provided by cloud services like Google Cloud Platform [7] as well as a number of niche companies and startups like H2O.ai. The promise of AutoML is yet to be seen at scale, but it represents an exciting opportunity to rapidly build and prototype a baseline machine learning or deep learning model for your use case and fast-track model development and deployment lifecycle. (4) Model improvements Algorithmic and model-based improvements require greater technical expertise, intuition and understanding of the business use case. Given the limited supply of data scientists who combine all the above skills, it’s not common for most businesses to invest significant resources and allocate the necessary time and bandwidth for innovative machine learning and deep learning research and development. As most business use cases and organizational data ecosystems are unique, a one-size-fits-all strategy is often not feasible nor advisable. This necessitates the requirement for original work to adapt existing or related applications to fit the businesses’ particular needs. Model improvements can come from distinct sources:
(5) Case Study: from BERT to RoBERTa In this section, I will describe a case study in large-scale model improvement for a state-of-the-art deep learning model for natural language processing. BERT, developed in 2018 by Google [8], has become the de-facto deep learning model to use for a range of NLP applications and has accelerated NLP research and use cases across the board. It yielded state-of-the-art performance on benchmarks like GLUE, which evaluate models on a range of tasks that simulate human language understanding. However, BERT’s tenure at the top of the GLUE leaderboard was soon replaced by RoBERTa, developed by Facebook AI, which was fundamentally an exercise in optimizing the BERT model further, as evidenced by its full name – Robustly Optimized BERT PreTraining Approach [9]. RoBERTA surpassed BERT in terms of performance on the basis of simple modifications including training the model for more epochs, feeding more data to the model, training the model on different data (longer sequences) with bigger batch size, and optimizing the model and design choices. These simple model improvement techniques increased the model score on the GLUE benchmark from 80.5% for BERT to 88.5% for RoBERTa, a highly significant outcome. Data In earlier sections, I discussed hyperparameter optimization and select model improvement strategies. In this section, I will describe the importance of focusing on the data to improve the performance of machine learning and deep learning models. In business, more often than not, improving the quality and quantity of training data yields stronger model performance. There are several techniques for a data-centric approach to machine learning and deep learning model improvement. (1) Data Augmentation The lack of gold standard annotated training data is a common bottleneck for developing and improving large-scale supervised machine learning and deep learning models. The cost of annotation in terms of time, expense, and subject matter expertise is a limiting factor to create massive labeled training datasets. More often than not, machine learning models suffer from overfitting and their performance can be improved by using more training data. Data augmentation techniques can be leveraged to expand the training dataset in a scalable fashion. The choice of data augmentation techniques depends on the kind of data. For instance, synthetic time-series data can be created by sampling from a generative model or probability distribution that is similar in summary statistics to the observed data. Images can be augmented by altering image characteristics like brightness, color, hue, orientation, cropping, etc. Text can be augmented by a number of methods including regex patterns, templates, substitution by synonyms and antonyms, backtranslation, paraphrase generation, or using a language model to generate text. Audio data can be augmented by modifying fundamental acoustic attributes like pitch, timbre, loudness, spatial location, and other spectrotemporal features. For specific applications, pretrained models can also be used to expand the original training dataset. In earlier sections, I discussed hyperparameter optimization and select model improvement strategies. In this section, I will describe the importance of focusing on the data to improve the performance of machine learning and deep learning models. In business, more often than not, improving the quality and quantity of training data yields stronger model performance. There are several techniques for a data-centric approach to machine learning and deep learning model improvement. (1) Data Augmentation The lack of gold standard annotated training data is a common bottleneck for developing and improving large-scale supervised machine learning and deep learning models. The cost of annotation in terms of time, expense, and subject matter expertise is a limiting factor to create massive labeled training datasets. More often than not, machine learning models suffer from overfitting and their performance can be improved by using more training data. Data augmentation techniques can be leveraged to expand the training dataset in a scalable fashion. The choice of data augmentation techniques depends on the kind of data. For instance, synthetic time-series data can be created by sampling from a generative model or probability distribution that is similar in summary statistics to the observed data. Images can be augmented by altering image characteristics like brightness, color, hue, orientation, cropping, etc. Text can be augmented by a number of methods including regex patterns, templates, substitution by synonyms and antonyms, backtranslation, paraphrase generation, or using a language model to generate text. Audio data can be augmented by modifying fundamental acoustic attributes like pitch, timbre, loudness, spatial location, and other spectrotemporal features. For specific applications, pretrained models can also be used to expand the original training dataset. Recent methods based on weak supervision, semi-supervised learning, student-teacher learning, and self-supervised learning can also be leveraged to generate training data with noisy labels. These methods are based on the premise that augmenting gold standard labeled data with unlabeled or noisy labeled data provides a significant lift in model performance. It’s now possible to leverage a combination of rule-based and model-based data augmentation techniques that can be engineered at scale using data augmentation platforms like Snorkel [10]. Another common scenario where models underperform is in the context of imbalanced data across categories of interest. In such scenarios with skewed data distribution, upsampling and downsampling of data and techniques like SMOTE are helpful in correcting the modeling results. The concept of having a training dataset, validation dataset, and test dataset is common in machine learning research. Cross-validation helps in shuffling the exact composition of these three datasets so that statistically robust inference can be made about the model performance. While classical approaches focus on three datasets with a single validation dataset, it’s good to have two different validation datasets, one drawn from the same distribution as the training data and the other drawn from the same distribution as the test data. This way you can better diagnose bias-variance tradeoff and use the right set of model improvement strategies as described above. (2) Feature engineering & selection Typical machine learning models are trained on data with numerous features. Another common technique to improve machine learning models is to engineer new features and select an optimal set of features that better improve model performance. Feature engineering requires significant domain expertise to devise new features that capture aspects of the complex nonlinear function that the machine learning model is learning to approximate. So, this method is not always feasible if the baseline model already captures a diverse set of features. Feature selection via programmatic approaches can help remove some correlated or redundant features that don’t contribute much to model performance. Methods to iteratively build and evaluate a model with a progressively increasing set of features, or iteratively reducing one feature at a time from a model trained with the entire set of features, help in identifying robust features. (3) Active learning Analysis of model errors can shed light on the kind of mistakes that the machine learning model makes. Reviewing these errors helps understand whether there are any characteristic patterns that can be addressed by some of the techniques described above. Additionally, active learning methods that focus on model mistakes that are closer to the decision boundary can provide a significant boost in performance once the model is already in production. In active learning, the new examples that the model is confused about and predicts incorrectly are sent for annotation to domain experts who provide the correct labels. This dataset that is reviewed and annotated by experts is incorporated back into the training dataset to help the retrained model learn from its previous errors. Conclusion Machine learning and deep learning modeling requires significant subject matter expertise, access to high-quality labeled data, as well as computational resources for continuous model training and refinement. Improving machine learning models is an art that can be perfected by systematically addressing the deficiencies of the current model. In this article, I have reviewed a set of methods focused on models, their hyperparameters, and the underlying data to improve and update models to attain the required performance levels for successful deployment. References
|
Archives
September 2024
Categories
All
Copyright © 2024, Sundeep Teki
All rights reserved. No part of these articles may be reproduced, distributed, or transmitted in any form or by any means, including electronic or mechanical methods, without the prior written permission of the author. Disclaimer
This is a personal blog. Any views or opinions represented in this blog are personal and belong solely to the blog owner and do not represent those of people, institutions or organizations that the owner may or may not be associated with in professional or personal capacity, unless explicitly stated. |
RSS Feed