BehaveRec sets 20-interaction threshold for long-term history gains in sequential recommendation

BehaveRec combines a causal session transformer, a 90-day bidirectional User History Encoder, and GraphSAGE on behavioral similarity graphs. A learned gate fuses the three signals per user. On the industrial dataset the model improves retrieval at rank 5 while a larger-parameter SASRec baseline still trails by 2.5%. Stratification by history length shows a monotonic dose-response: negligible lift below 10 events, rising to +0.017 R@5 for users above 101 events. The pattern holds across three public datasets with varying history density.

The 20-interaction deployment rule directly affects short-form video feeds that millions encounter daily. Viewers with sparse early sessions receive little benefit from added history or community signals, so platforms can default to lightweight session models and reserve heavier encoders for richer profiles. This avoids unnecessary compute while concentrating personalization where behavioral signals stabilize intent prediction. The category-level result leaves item-level vocabulary performance as an open question requiring further runs.

Related work on SASRec and GraphSAGE established the base components; BehaveRec adds the per-user gate and explicit history-length stratification absent from those papers. Prior sequential models rarely reported user-level breakdowns, masking the point at which long-term signals become reliable. The current study therefore supplies the missing operational cutoff rather than assuming universal benefit from added context.

Next steps include multi-run item-level evaluation and production A/B tests that track watch-time lift above the 20-interaction threshold. If confirmed, the rule offers a simple pre-filter for routing users to different model stacks without retraining the entire system.