A Multi-behavior Recommendation Based on Disentangled Graph Convolutional Networks and Contrastive Learning

Traditional recommendation models typically rely on a single type of user-item interaction data, which presents a serious challenge due to data sparsity. Multi-behavior recommendation models leverage various available user behaviors in the recommendation scenario as auxiliary data to assist in predicting user-item interaction. However, existing multi-behavior recommendation models do not take into account potential factors that affect multi-behavior interaction, or the differences that exist between different types of behaviors. In this paper, we propose a Multi-Behavior Recommendation Based on Disentangled Graph Convolutional Networks and Contrastive Learning (DCMBR). Specifically, we construct subgraphs for new dissatisfied behavior, and use disentangled convolution networks to separate potential factors that affect interaction between users, items, and behaviors, so as to reconstruct node features for users under different behaviors. Then, users’ multi-behavior characteristics are aggregated using contrastive learning to achieve personalized multi-behavior information aggregation. Experimental results on two datasets demonstrate that DCMBR can effectively leverage multi-behavioral data, and significantly improve recommendation performance compared to the optimal baseline.