LRGCN: Linear Residual Graph Convolution Network Recommendation System

Abstract

With the vigorous development of the Internet and the continuous expansion of the scale of product information, people put forward higher requirements for filtering redundant information in product recommendation. Researchers use a bipartite graph to model the interaction between users and items so that Graph Convolutional Network (GCN), the most advanced graph representation model, can be widely and successfully applied in the recommendation system. However, GCN can cause the representation of nodes over smooth and the layers of most GCN models cannot be stacked deep to capture higher-order cooperative signals. In this work, we study the recommendation system by optimizing the over smoothing effect of GCN. Firstly, we remove the nonlinear part of GCN in the message passing process. Secondly, we introduce the residual network structure and propose the Linear Residual Graph Convolution Network (LRGCN) network model so that the number of stacked layers can be effectively increased while maintaining good performance. Finally, we optimize the negative sampling strategy to improve the performance of the recommendation system by 2.8%. Our proposed model is linear and achieves better results on three different real data sets than the baselines.