GCMCSR: A New Graph Convolution Matrix Complete Method with Side-Information Reconstruction

Kun Niu, Yicong Yu, Xipeng Cao, Chao Wang

Abstract: In this work, we propose a novel Graph Convolutional Matrix Completion with Side-information Reconstruction (GCMCSR) model for the recommender system. For most recommender systems, the side-informations of users and items are usually utilized as the input of the model. However, when new users or new projects are included, the system's performance can degrade significantly. In GCMCSR, to solve this problem, we take the side-information as labels to predict under a multi-task learning framework, which contains a graph-based matrix completion task and a side-information reconstruction task. We borrow the idea of Graph Convolutional Matrix Completion (GCMC) to acquire user/item representation by spatial information extracted from the user-item bipartite graph. The experiment results show that our model achieved state-of-the-art performance on all three public datasets.