Spurious Correlation Knowledge Graph Disentanglement for Multi-behavior Recommendation

Tongxin Xu, Chenzhong Bin, Cihan Xiao, Zhixin Zeng, Yunhui Li

Abstract: Multi-behavior recommendation systems have been shown to effectively mitigate data sparsity in target behavior by leveraging various auxiliary behaviors. Existing multi-behavior recommendation methods primarily focus on modeling auxiliary behaviors and applying them to target behavior recommendations. However, the interaction correlations between item features and user intents in the auxiliary behaviors contain both genuine and spurious correlations. Indiscriminately modeling these correlations can easily lead to a degradation in recommendation performance. Therefore, we propose a novel SCKGD framework for multi-behavior recommendation. In particular, we construct a spurious correlation knowledge graph (SCKG) from the item perspective to capture the spurious correlations between items. Based on the generated SCKG, a time-projected knowledge encoder is designed to dynamically weight the strength of spurious correlations between items and learn the corresponding item embeddings. To disentangle the spurious correlations, we propose a disentangling contrastive learning method that preserves the genuine correlated semantics aligning with the user’s target intent. Concurrently, to enrich the supervision signal in the target behavior, we perform inter-behavior contrastive learning to transfer the genuine correlated semantics from auxiliary behaviors into the target behavior. Experiments on three datasets demonstrate that our framework outperforms state-of-the-art (SOTA) methods and disentangles spurious correlations. Our implementation is available (https://github.com/LokHsu/SCKGD).

External IDs:doi:10.1007/978-981-92-0363-5_22