Negative Samples Selection Can Improve Graph Contrastive Learning in Collaborative Filtering

Nowadays, graph collaborative filtering is widely used in recommender system as an important technique. In order to alleviate the common data sparsity problem in collaborative filtering, contrastive learning techniques have been utilized to assist the modeling of user-item interaction graphs. However, the existing graph contrastive learning methods in collaborative filtering mainly focus on the construction of the augmented view and the loss function, which ignores the negative samples selection. They treat all other nodes as negative samples in learning process, which may be unreasonable. In fact, if two users have extremely similar preferences, they should not be treated as negative samples of each other. To address the aforementioned issue, we investigate various techniques for negative samples selection and introduce a graph contrastive learning framework called NSSGCL (Negative Samples Selection Graph Contrast Learning), which consists of two parts. First, we measure the similarity of nodes in terms of both their structure and semantics, and consider nodes that are highly similar in both structure and semantics aspects to be potential positive samples of each other. Second, we propose a debiasing algorithm that utilizes the probability distribution of positive and negative samples to further estimate the actual contribution of the potential positive samples obtained above. A substantial number of experiments on real datasets demonstrate the effectiveness of our strategy, with our approach achieving a 12% and 14% improvement compared to baseline methods on the ML-1M and Yelp datasets, respectively.