RESEARCH · RESEARCH · #2706
BehaveRec paper studies when long-term behavioral history helps sequential recommendation
Authors introduce BehaveRec, a three-signal sequential-recommendation framework that combines a causal session transformer, a bidirectional 90-day User History Encoder, and a GraphSAGE behavioral user graph with a learned gate. On an industrial dataset (137M sequences, 12.6M users) BehaveRec improves R@5 by +3.2% over SASRec (and +2.5% versus a larger-capacity SASRec baseline), with benefits negligible below ~10 past interactions and rising monotonically with history richness (reaching +0.017 R@5 for 101+ events); the paper derives a practical deployment rule of roughly 20 interactions per user and notes item-level results are directionally mixed and left open.
KEY POINTS
- Authors introduce BehaveRec, a three-signal sequential-recommendation framework that combines a causal session transformer, a bidirectional 90-day User History Encoder, and a GraphSAGE behavioral user graph with a learned gate.
- On an industrial dataset (137M sequences, 12.6M users) BehaveRec improves R@5 by +3.2% over SASRec (and +2.5% versus a larger-capacity SASRec baseline), with benefits negligible below ~10 past interactions and rising monotonically with history richness (reaching +0.017 R@5 for 101+ events); the paper derives a practical deployment rule of roughly 20 interactions per user and notes item-level results are directionally mixed and left open.
- Provides an empirical rule and stratified evidence for when adding long-term history and behavioral-graph signals yields meaningful gains in sequential recommendation, helping guide deployment trade-offs.
WHY IT MATTERS
Provides an empirical rule and stratified evidence for when adding long-term history and behavioral-graph signals yields meaningful gains in sequential recommendation, helping guide deployment trade-offs.