Glossary · E-commerce ML

Real-Time Personalization

also: session-based recommendation · in-session personalization

Definition

Real-time personalization adapts product recommendations, content, and pricing within a session based on immediate behavior signals — dwell time, scroll depth, added items, search queries. Contextual-bandit systems with streaming feature stores enable policy updates in milliseconds, producing 10–25% lift over batch-trained models.

Traditional recommenders train nightly on batched user-item interactions. Real-time personalization closes the loop: session-level signals update features in a feature store, which feeds a bandit or online-learning model that selects the next action within milliseconds. The architecture requires low-latency feature serving (Redis, Feast), an online policy (contextual bandit, reinforcement learning), and logged-experience training that handles selection bias. Gains compound with session depth — later-session decisions benefit more.

Essays on this concept