Graph Contrastive Learning with Positional Representation for Recommendation

Abstract

Recently, graph neural networks have become the state-of-the-art in collaborative filtering, since the interactions between users and items essentially have a graph structure. However, a major issue with the user-item interaction graph in recommendation is the absence of the positional information of users/items, which limits the expressive power of graph recommenders in distinguishing the users/items with the same neighbours after propagating several graph convolution layers. Such a phenomenon further induces the well-known over-smoothing problem. We hypothesise that we can obtain a more expressive graph recommender through graph positional encoding (e.g., Laplacian eigenvector) thereby also alleviating the over-smoothing problem. Hence, we propose a novel model named Positional Graph Contrastive Learning (PGCL) for top-K recommendation, which aims to explicitly enhance graph representation learning with graph positional encoding in a contrastive learning manner. We show that concatenating the learned graph positional encoding and the pre-existing users/items’ features in each feature propagation layer can achieve significant effectiveness gains. To further have sufficient representation learning from the graph positional encoding, we use contrastive learning to jointly learn the correlation between the pre-exiting users/items’ features and the positional information. Our extensive experiments conducted on three benchmark datasets demonstrate the superiority of our proposed PGCL model over existing state-of-the-art graph-based recommendation approaches in terms of both effectiveness and alleviating the over-smoothing problem.