Cascading Hypergraph Convolution Networks for Multi-Behavior Sequential Recommendation

Dan Lu, Hao Zhang, Lijie Li, Shiqing Wu, Guandong Xu

Abstract: In current research on recommendation systems, exploring multi-behavior sequence recommendation has become a crucial topic. It is well known that user interactions on online platforms (such as social media websites, news aggregation applications) involve not only singular actions like reading or clicking but also various other behaviors such as commenting, sharing, and bookmarking. These diverse interactions reflect different facets of user preferences in item interaction sequences. Therefore, understanding and integrating these diverse behaviors to effectively represent user preferences personalized is essential. To better utilize temporal information, we propose a Cascading Hypergraph Convolution Networks For Multi-Behavior Sequential Recommendation (CHMSR) framework. CHMSR aims to capture the underlying behavioral preferences in user sequence interactions and predict future user behaviors through cascading neural networks. Specifically, CHMSR first independently encodes each behavior sequence to extract user interests from the complex behavior relationships in the sequences. Then, it aggregates user preferences for different items through hypergraph convolution to capture global preferences. Furthermore, we employ cascading neural networks in behavior chains to capture directional feature dependencies in multi-behavior sequences, propagating messages from upstream to downstream behaviors. In this way, CHMSR can comprehensively consider user interests in multi-behavior sequences and global preferences for items, thus providing more accurate and personalized recommendations. Our experimental results demonstrate that CHMSR significantly outperforms existing recommendation methods on various datasets, validating its effectiveness and practicality