Negative Samples Selection Can Improve Graph Contrastive Learning in Collaborative Filtering

Yifan Shao, Xu Cai, Fangming Gu, Ximing Li

Abstract: 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.