Fast Sub-ML Usecase Development using Feature Stores

Speaker: Achint Thomas, Scribble Data

Widespread adoption of machine learning (ML) in industry is still a challenge today due to resource constraints and justifying cost vs. outcomes. ML approaches require high skill and rely on large volumes of data. In this talk, we argue for Sub-ML - a class of applications simpler than traditional ML approaches and often designed to be used in decision support systems. Characterized by their speed in realizing business value and support for diverse use cases, Sub-ML applications still require guarantees of correctness, transparency, and auditability in the data transformation process. We draw on our experience in the fin-tech, ed-tech and e-commerce domains to lay out design choices for feature stores to enable Sub-ML, including constraining the problem space, bundling capabilities for fast development, and incorporating a data consumption layer.