Advancing Cascading Residual Graph Convolution Networks for Multi-behavior Recommendation: An Innovative Approach Within Representation Learning

Auxiliary behavior data is introduced into multi-behavior recommendation (MB-Rec) to alleviate the recommendation data sparsity. Recent studies have shown that different auxiliary behaviors usually appear in a specific order (e.g., view > cart > buy), and the latter behavior interactions usually exhibit stronger user preference signals than the former ones. The use of cascading graph convolution networks to capture these dependencies is an effective approach. However, in this structure, preference learning performed on target behaviors is susceptible to the quality of the upstream representation. In this work, we propose an Enhanced Cascading Residual Graph Convolution Networks (ECR-GCN) for MB-Rec. Specifically, we design an auxiliary behavior representation enhancement layer to optimize the residual embeddings between behaviors against biased and low-quality embeddings coming from upstream behavior data, and to enhance the effect of preference representation learning on target behaviors. ECR-GCN outperforms the state-of-the-art models on two datasets by an average of 73.52% and 53.18% in terms of HR@10 and NDCG@10, respectively. The model source code is available at: https://github.com/lhybq/ECR-GCN .