Declarative Feature Engineering at PayPal

PayPal describes their declarative (config-based) feature engineering paradigm used for real-time fraud detection ML across 400 million users.

• •Features are declared as "what" they look like rather than "how" to construct them, abstracting execution platform details from data scientists
• •Features are split into three complexity levels: simple features (roll-up aggregations, categorical), code-based features, and analytical platform features (graph, NLP, anomaly detection)
• •Simple features are handled via a self-service UI/API allowing data scientists to simulate and productize features without direct engineer support
• •The system auto-generates optimized data pipelines, enforces Time Travel for Point-in-Time historical backfill, and registers features to a Feature Store for cross-team reuse
• •The approach reduces Time to Market (TTM) and Total Cost of Ownership (TCO) by detecting duplicate features and recommending reuse before production