Improving Graph Collaborative Filtering from the Perspective of User–Item Interaction Directly Using Contrastive Learning

Abstract

Graph contrastive learning has demonstrated significant superiority for collaborative filtering. These methods typically use augmentation technology to generate contrastive views, and then train graph neural networks with contrastive learning as an auxiliary task. Although these methods are very effective, they do not consider using contrastive learning from the perspective of user–item interaction. As a result, they do not fully leverage the potential of contrastive learning. It is well known that contrastive learning can maximize the consistency of positive pairs and minimize the agreement of negative pairs. Collaborative filtering expects high consistency between users and the items they like and low consistency between users and the items they dislike. If we treat the items that users like as positive examples and the items they dislike as negative examples, contrastive learning can work very well with the goal of collaborative filtering. Based on the above understanding, we propose a new objective function called DCL loss, which improves graph collaborative filtering from the perspective of user–item interaction when Directly using Contrastive Learning. Extensive experiments have shown that when a model adopts DCL loss as its objective function, both its recommendation performance and training efficiency exhibit significant improvements.