Adaptive Augmentation and Neighbor Contrastive Learning for Multi-Behavior Recommendation

Graph-based contrastive learning has achieved significant success in multi-behavior recommendation systems. However, two challenges still persist: (1) Edge weights are neglected when constructing augmented views based on the interaction graph. It leads to the omission of crucial nodes or edges during the augmentation process. (2) During contrastive learning, positive pairs are adopted by using topology structure. However, we argue that semantic similarity neighbors are the better positive samples. To address these challenges, we propose corresponding solutions: (1) We construct the weighted conversion graph and convert the interaction graph into a weighted adjacency matrix, then utilize Bernoulli distribution for masking operations to achieve adaptive augmentation. (2) We apply clustering algorithms to make semantic similarity nodes into same cluster and retain the neighbors with node centrality values within specific range in each cluster, then select appropriate positive pairs during contrastive learning. Considering these two aspects, we propose the Adaptive Augmentation and Neighbor Contrastive Learning (AANCL) framework for multi-behavior recommendation. The evaluation results on three real-world datasets validate the effectiveness of the proposed model.