BASG: A Behavior-Augmented Semantic Generalization Framework for Zero-Shot Cross-Domain Sequential Recommendation

Abstract

Training sequential recommenders for interaction-disjoint target domains is challenging: no target interaction data are available for training, model selection, or calibration. Semantic item representations support transfer across non-overlapping catalogs but do not encode query-local temporal signals or interaction-derived properties such as popularity and co-occurrence. We propose BASG (Behavior-Augmented Semantic Generalization), a backbone-compatible framework combining a source-trained semantic scorer with a separately trained Behavior Expert via a common candidate-score interface. The expert integrates query-local temporal features with popularity- and co-occurrence-related proxies predicted from frozen item-text embeddings by source-fitted mappers. Static Fusion standardizes and combines the two score vectors using a source-selected coefficient fixed during target evaluation. Catalog-level text representations and transferred property proxies are precomputed offline and reused at inference, without target-side calibration or aggregate target-domain statistics. Across three semantic backbones and three target domains including the cross-platform Steam dataset, BASG improves mean NDCG@10 by 0.43–1.68 percentage points in all nine backbone–target settings. In SAGERec-based analyses on the two Amazon targets, the interaction-property channel shows the larger removal effect, while source-calibrated Static Fusion transfers more reliably than the tested source-trained routers. These results support BASG as a strategy for integrating transferable behavioral evidence under a strict zero-shot protocol.