Contrastive Multi-View Self-Supervised Learning for Heterogeneous Information Network

Self-supervised learning constructs supervised signals inside samples without relying on external labels, which is becoming a promising research direction. Recently, works on self-supervised learning by maximizing local-global mutual information on networks have achieved state-of-the-art performance comparable to semi-supervised graph neural networks (GNNs). However, these methods have not explored the collaborative relationship of multiple meta-path views, and the global representation is weakened by irrelevant nodes which participate in the average operation over all nodes. In this paper, a self-supervised approach based on mutual information for heterogeneous information network embedding is proposed. Specifically, it utilizes the contrast of multiple meta-path views to supervise each other, and positive samples are selected to obtain a robust global representation. Experimental results demonstrate the proposed method has competitive performance over the existing mutual-information-based ones and even outperforms some supervised learning methods.