Weighted Graph Convolutional Network for Collaborative Filtering Considering Entity Similarity

Abstract

The core idea of collaborative filtering is to use historical interaction data of users and items to find similar users or items for recommendation. Most of the recommendation work based on the graph field is to directly use the Graph Convolutional Network (GCN) on the user-item bipartite graph. This kind of graph cannot reflect the similarity of information between the homogenous nodes of the users (items). To solve this problem, a Weighted Graph Convolutional Collaborative Filtering recommendation model (WGCCF) considering entity similarity is proposed. In this work, the user-item bipartite graph will be transformed into two homogeneous graphs. Firstly, the user-item scoring matrix is decomposed by singular value decomposition, and the vector representation of users(items) is obtained to calculate the similarity between users (items). Then the similarity is used as the edge weight of nodes in user(item) homogeneous networks and a weighted network based on node similarity is constructed. Finally, graph convolution operation is performed on the user (item) graph to extract the hidden layer representation of the user (item), and the vector inner product combination method is used to predict the score. Experimental results show that compared with the existing recommendation algorithms for matrix factorization and graph convolution, the index performance of RMSE and MAE can be improved up to 12.72% and 15.99%.