Graph neural network recommendation algorithm based on self-supervised learning

Abstract

Researchers have developed a specific sub-class of deep learning methods, the so-called Graph Neural Network (GNN), to perform inference on data represented by graphs. It has gained increasing popularity in recent years due to numerous domains utilizing graph-structured data such as e-commerce, social networks, and biomedical science. Despite their effectiveness, the current supervised learning methods applied in GNN have led to several weaknesses due to their heavy data reliance. Self-supervised learning (SSL), a promising improvement over (semi-)supervised learning, eliminates the need for manual annotation. By learning better item features’ latent relationships, SSL overcomes the label sparsity problem by improving item representation learning and improving generalization. This project explored various designs of existing SSL algorithms on three benchmark graph datasets: Gowalla, Yelp2018, and Amazon Book. We first conducted a review of existing SSL methods for graph learning. From that, we proposed our GNN architecture using LightGCN [7] with contrastive learning - a type of SSL, and then present the results of our experiments in the recommender system application. Empirical results obtained from this project showed that LightGCN with SSL can improve performance metrics.