Self-restrained contrastive enhanced network for graph structure learning

Abstract

Existing graph neural networks (GNNs) are mostly applied to representation scenes with complete graph structure. However, the graph structures of complex systems from the real world are generally missing or misconnected. In addition, due to the indistinguishability of graph node features, most of the existing GNNs are limited by the over-smoothing problem when performing classification tasks. To address these issues, we propose a model for learning and optimizing graph structure—Self-restrained Graph Contrastive Enhanced Network (GCEN). Firstly, the multiple branch generation structure is adopted to obtain various graph structures, which are updated by multi-order information fusion to obtain a new graph structure. Secondly, the performance of each generated structure branch is improved by self-restrained network, and then the structure information of the graph is continuously optimized to obtain a high-quality graph structure. Finally, the feature contrast enhancement module is utilized to adaptively adjust the features of nodes to obtain distinguishable features, thus alleviating the over-smoothing problem of GNNs in classification tasks. Extensive experiments on several benchmark datasets indicate that our GCEN is superior to several models for learning task-specific graph structure in optimizing the graph structure.