Abstract
Multi-behavior recommendation (MBR) alleviates target behavior sparsity by leveraging abundant auxiliary behaviors. However, most MBR methods are trapped in a dilemma. Separately aggregating multiple behaviors struggles to fully exploit cross-behavior interaction information, while collectively aggregating different behaviors into a unified graph may cause the target behavior to be submerged among massive auxiliary behaviors. To this end, we propose the Semantic-struCtural crOss-calibRation modEl (SCORE), a structure-conditioned calibration framework for MBR. SCORE adopts a decouple–enhance–calibrate strategy to fully exploit cross-behavior interactions while robustly reconstructing target-related user representations. Specifically, SCORE first decouples target semantics from unified representations through a shared modulation space under joint multi-behavior supervision. It then enhances unified representations through structural refinement on the target-behavior graph, reconstructing target-consistent structural context. Finally, SCORE reconciles semantic and structural references through structure-conditioned gating to obtain calibrated representations. Extensive experiments on three real-world benchmarks demonstrate that SCORE greatly improves over thirteen competitive baselines by up to 61.73% on HR@10 and 46.04% on NDCG@10.
Funding Information
  • National Natural Science Foundation of China (62432003)
  • National Natural Science Foundation of China (U25A20431)

This publication has 26 references indexed in Scilit: