Improving Graph Convolutional Networks with Transformer Layer in social-based items recommendation

Abstract

With the emergence of online social networks, social-based items recommendation has become a popular research direction. Recently, Graph Convolutional Networks have shown promising results by modeling the information diffusion process in graphs. It provides a unified framework for graph embedding that can leverage both the social graph structure and node features information. In this paper, we improve the embedding output of the graph-based convolution layer by adding a number of transformer layers. The transformer layers with attention architecture help discover frequent patterns in the embedding space which increase the predictive power of the model in the downstream tasks. Our approach is tested on two social-based items recommendation datasets, Ciao and Epinions and our model outperforms other graph-based recommendation baselines.