Disentangling Multi Behaviors for Recommendation Based on Graph Contrastive Learning

Abstract

Recommendation systems play a crucial role in efficiently filtering vast amounts of information, and supporting businesses across various domains such as e-commerce, social media, and online entertainment. Traditional recommendation methods, including collaborative filtering, often suffer from data sparsity issues, limiting their effectiveness. Multi-behavior recommendation has emerged as a promising solution by leveraging implicit user interactions, such as clicks and favorites, to improve recommendation quality. Existing graph-based multi-behavior recommendation models typically construct a single interaction graph, leading to increased computational costs and potential interference between different behavioral signals. To address this, we propose a disentangled multi-behavior recommendation model that explicitly separates user behavior graphs, propagates user signals independently, and refines representation learning through contrastive learning. This approach enables better capture of user pReferences while mitigating the negative effects of data sparsity. Additionally, we introduce distance correlation coefficients to regulate the embedding representations of different behaviors, preventing cross-interference and further improving recommendation accuracy. Experimental results on real-world datasets demonstrate that our model outperforms existing single- and multi-behavior recommendation methods. Ablation studies verify the effectiveness of different components, and further experiments highlight the model’s ability to disentangle user behaviors, offering a more interpretable and effective recommendation framework.