Quaternary converter based balanced graph contrastive learning for recommendation

Abstract

Graph contrastive learning (GCL) has emerged as a highly effective collaborative filtering method for recommender systems in recent years. However, existing collaborative filtering methods based on GCL often excessively prioritize user-side information, leading to inadequate exploration of user–item information. Furthermore, these methods generate contrastive views through data augmentation, which is prone to noise interference. To address these issues, we propose a balanced graph contrastive learning framework (BGCL). Specifically, BGCL incorporates a quaternary converter that introduces negative user based on triples (user, positive item, negative item), to provide the GCL module with embeddings that treat users and items balancedly. Subsequently, BGCL includes a noiseless GCL module that conducts contrastive learning on the embeddings after approximately infinite layers of convolution and the embeddings after k-layer graph convolutional networks to mitigate noise interference. We conducted experiments comparing our algorithm with 15 alternative approaches using real-world datasets, and the results demonstrate that our algorithm outperforms state-of-the-art methods in terms of recommendation accuracy and convergence speed.