Enhanced contrastive learning with multi-aspect information for recommender systems

Abstract

In recent years, graph neural networks (GNNs), as the state-of-the-art technology have advanced the solutions of recommender systems (RSs). However, the representations of nodes obtained from GNN-based models are always suboptimal when facing sparse interaction data. Contrastive learning on graphs has been proven to be effective in alleviating the influence of data sparsity. Most related works perform contrastive learning through self-discrimination. As there is extensive homophily among nodes in RSs, we consider it irrational to ignore the self-supervised signal from neighboring nodes, which is crucial for optimizing the embedding representation of nodes. To address these issues, we propose a model named enhanced contrastive learning with multiaspect information (ECMI) for RSs, which aims to capture the self-supervised signals from both the nodes themselves and their neighbors. Specifically, we conducted contrastive learning in a two-layer space separately. We consider structurally connected homophilic node pairs as positive samples in explicit space while introducing a social relationship network and an item entity network in implicit space. These two networks can be combined with the interaction network to conduct a two-channel training procedure. This procedure can build more effective contrastive learning by exploring the similar potential neighbors of nodes. We conduct experiments on four public datasets, and the results consistently validate the effectiveness of the ECMI.