Towards efficient graph neural networks for large-scale recommender systems

Abstract

In the era of information explosion, recommender systems are vital tools for delivering personalized recommendations for users by forecasting their future behaviours based on historical user-item interactions. To model these interaction behaviours, Graph Neural Networks (GNNs) have remarkably boosted the prediction performance of recommender systems due to their strong expressive power of capturing high-order information in user-item interactions through multi-layer embedding propagations. Nonetheless, classic Matrix Factorization (MF) and Deep Neural Network (DNN) approaches still play an important role in real-world, large-scale recommender systems due to their scalability advantages. Despite the existence of acceleration solutions, it remains an open question that whether GNN-based recommender systems can scale as efficiently as classic MF and DNN methods. In this thesis, we systematically reviewed the existing GNN-based recommendation models and investigated the inefficiency of these prior works. In response to this limitation, we propose a Linear-Time GNN (LTGNN) to scale up GNN-based recommender systems to achieve comparable scalability as the classic and efficient Matrix Factorization approaches while maintaining the powerful expressiveness for superior prediction accuracy. Detailed theoretical analysis are conducted to illustrate the expressive capability and training efficiency of the proposed LTGNN. We also present extensive empirical experiments and ablation studies to validate and understand the effectiveness and scalability of the proposed algorithm.