AI

BehaveRec: When Does Long-Term Behavioral History Help Sequential Recommendation? A Study with Session, History, and Community Signals

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 28/10/2026, 04:03:25 UTC.

Briefings

A study titled BehaveRec investigates when long-term behavioral history improves sequential recommendation. Using an industrial dataset of 137M sequences from 12.6M users, the research found that BehaveRec improves R@5 by +3.2% over a baseline, with a user-level stratification showing a clear dose-response relationship where advantages grow with history richness, reaching +0.017 R@5 for users with 101+ events. The study concludes that a practical deployment threshold of roughly 20 interactions per user is needed to justify the added complexity, while noting that a preliminary single-run check on the full item-level vocabulary yielded mixed results.

Key points

• BehaveRec improves R@5 by +3.2% over a baseline model using a three-signal framework combining a causal session transformer, a bidirectional User History Encoder over a 90-day window, and a GraphSAGE module over a user-user graph.
• User-level stratification reveals a clear dose-response relationship where BehaveRec's advantage is negligible below roughly 10 interactions and grows monotonically with history richness, reaching +0.017 R@5 for users with 101+ events.

Context

The study addresses sequential recommendation models that typically predict next actions from recent session interactions, which limits personalization when short-term signals are weak. It proposes BehaveRec as a three-signal framework to determine when incorporating longer-term history and community signals is beneficial, using an industrial dataset of 137M sequences from 12.6M users.

Limitations

The results are established for category-level intent prediction; a preliminary single-run check on the full item-level vocabulary yielded mixed results and is scoped as an open question. The study uses a 90-day window for history and a single GPU for training, with a preliminary check on the full item-level vocabulary yielding mixed results.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 28/10/2026, 05:01:11 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeig7niugxocbknfe56kbfgpjhzcz3brepvhksxrkqcr7nufdlt2yua