High-Order Subgraph Convolution Networks with Auxiliary Self-Supervised Task for Recommendation

In recent years, Graph Convolution Networks have achieved promising results on recommendation tasks. Starting from the influential Pinsage, many other works have been proposed and techniques such as Light Graph Convolutional Networks, self-supervised Graph Learning, and Interest-aware Message-Passing have gained popularity thanks to their good performance. While self-supervised Graph Learning explores the advantages of learning a so-called auxiliary self supervised tasks, Interest-aware Message-Passing finds its success by using high-order graph convolution inside subgraphs. The subgraph represents groups of similar users together with their item interactions. In this work, starting from these two novel techniques, we propose an approach that combines the high-order subgraph with an auxiliary self supervised task. Experimental results on benchmark datasets show that our technique achieves state-of-the-art results.