CateReg: category regularization of graph convolutional networks based collaborative filtering

Recent years have witnessed the compelling improvement of collaborative filtering (CF) techniques that enjoy the expressive power of graph convolutional networks (GCNs). Unfortunately, there is an intrinsic drawback that limits GCNs to benefit from deep architecture, namely, the oversmoothing phenomenon wherein all nodes in the graph asymptotically incline towards the same representation. We find that this issue also exists in graph neural network based CF models. To cope with this problem, in this work we propose a regularization technique termed CateReg that punishes the distances between two nodes in the embedding space according to the category(i.e., user-item, user-user and item-item interactions) of their relationships. By optimizing the calculation process, we reduce the time complexity of distance computing from o(n2) to o(n). The evaluation experiments are carried out over several representative recommendation datasets. The effectiveness of our regularization method is demonstrated with about 2.3% and 2.6% improvements on the state-of-the-art model w.r.t Recall and NDGC, respectively.