Multi-behavior contrastive learning with graph neural networks for recommendation

Traditional recommendations typically prioritize modeling the target user's one type of behavior while ignoring other auxiliary behaviors, resulting in low recommendation accuracy. Meanwhile, the exploration of multiple types of behavioral patterns promotes the performance of recommendation systems, but two challenges remain: capturing the complex dependencies among multi-type behaviors based on users' personalized preferences and dealing with the problem of low recommendation accuracy owing to sparse target behavioral supervision signals. To address these challenges, we propose a novel recommendation framework named Multi-Behavioral Contrastive Learning Recommendation (MBCLRec) based on graph neural networks and contrastive learning. Specifically, we first propose a behavioral context information encoder to encode information and facilitate message passing for cross-type behaviors. Then, we design a multi-behavior contrastive learning module to mine the similarities and differences between the target and auxiliary behaviors. This is achieved by selecting appropriate pairs of positive and negative samples and capturing the dependencies among cross-type behaviors through multi-behavioral contrastive learning and self-attention networks. We further model high-order multi- behavioral relationships using a cross-propagation layer interrelationship learning module through a multi-head self-attention network. Finally, we unify the user multi-behavioral and high-order multi-behavioral relationship information in the recommendation system, which can effectively alleviate data sparsity. Extensive experiments on three real-world datasets validate that MBCLRec has a significant advantage over various state-of-the-art baselines.