Social and Categorical Signals in Contrastive Learning for Recommendation Systems

Abstract

Social recommendation systems that use graph neural network (GNN) models are effective in addressing the data sparsity issue present in collaborative filtering models. Social homophily and item similarities are critical factors that affect users' preferences, and GNN models must capture these factors while incorporating users' interaction behaviors. In this work, we propose SCCL, a recommendation model that jointly captures social influence and item similarity signals with cross-view contrastive learning. We constructed a user-user social graph from social networks and item-item graphs from common tags. In our model, user-user relations are represented as a homogeneous graph, and item-item relations are represented as hypergraphs. We demonstrate the effectiveness of our model on two real-world datasets.