April 2023 Summaries

The shift from batch to real-time machine learning is becoming increasingly common as companies seek to improve personalization and stay ahead of the competition. Real-time ML refers to generating predictions online with at least an hourly frequency, distinguishing it from batch ML which generates predictions on a periodic, infrequent basis. The key to making this transition lies in understanding when and whether real-time ML makes sense for your product or use case, considering factors such as accuracy, cold-start problem, costs, and user expectations. A thorough assessment of the industry, technology stack, and business needs is necessary to determine if real-time ML is a suitable fit. With proper evaluation and planning, companies can leverage real-time features and models to improve their products' responsiveness and competitiveness.

FanDuel, a leading online gaming company, recently discussed its journey to implementing a production machine learning (ML) platform, called Omni, at an event in New York City. The company's engineering and data science teams adopted Tecton's feature platform to power their ML models and data-driven applications. FanDuel chose Tecton over building its own solution due to the need for managed solutions, cost-effectiveness, and seamless integration with their existing stack. Tecton provided a way for different teams to contribute to an inter-source repo, resulting in a growing library of features. The integration has enabled personalization efforts and responsible gaming models, processing hundreds of millions of rows of data and scoring millions of results. With Tecton's help, FanDuel has streamlined its feature management process and developed more accurate models, driving business growth and enhancing the customer experience.

The challenge of real-time machine learning lies in bridging the gap between the online and offline environments, where models are trained on historical data but need access to real-time feature values for predictions. This "online / offline skew" problem can lead to poor model quality if not addressed, as models may be trained on outdated or inaccurate features. To reduce skew, teams can use feature stores and platforms that address the different sources of skew, including differences in online and offline feature logic, historically accurate training data generation, and changes in real-world distributions. A well-designed feature platform can help alleviate these challenges, allowing teams to focus on delivering business value with their real-time machine learning systems.