BehaveRec: When Does Long-Term Behavioral History Help Sequential Recommendation? A Study with Session, History, and Community Signals
Summary
- Researchers proposed BehaveRec, a three-signal framework for sequential recommendation that combines causal session transformers, a 90-day User History Encoder, and a GraphSAGE module over user behavioral embeddings.
- Tested on an industrial dataset of 137M sequences, BehaveRec improves R@5 by 3.2% over SASRec while using only 5.67M parameters and training on a single GPU in under two days.
- User-level stratification shows a dose-response relationship, where performance gains grow monotonically with history richness starting around 10 interactions.
- The study establishes a practical deployment threshold of roughly 20 interactions per user, indicating session-only models remain preferable in sparse regimes.
Why it matters
Provides empirical guidelines for when incorporating costly long-term historical and graph signals actually improves recommendation systems versus when session-only models suffice.