Self-Supervised Graph Representation with Proximity Matrix

Abstract

Graph representation learning(GRL) aims to extract valid graph information to generate informative representations for handling various tasks on graphs. Supervised graph representation learning (SGRL) has achieved remarkable success recently. However, when labels are missing, or task-free representations are expected, self-supervised graph representation learning (SSGRL) is a promising paradigm. Graph auto-encoders (GAEs) are one of the prevalent frameworks in SSGRL. There are two issues that hinder the effectiveness of existing GAEs: a) There is no appropriate strategy to fuse attributive and structural information in the reconstruction target; b) The decoder cannot fully retain valid information for classification tasks. In this work, we propose a self-supervised graph auto-encoders via proximity matrix reconstruction (PMGAE) to address the above issues. For the proper information fusion, a proximity matrix is designed as the reconstruction target. Then, we design a cosine decoder to enhance the decoder's ability to retain information. Moreover, to adapt PMGAE to the large-scale graph, a random sub-matrix reconstruction mechanism is introduced to reduce memory consumption and running time in the decoding stage. In the experiment, we used seven benchmarks to verify the effectiveness of PMGAE. The results show that the proposed PMGAE is significantly improved compared to the state-of-the-art methods.

Keywords: Graph Representation Learning Graph Auto-Encoders, Self-supervised learning