Self-Supervised Graph Learning with Segmented Graph Channels

Abstract

Self-supervised graph learning adopts self-defined signals as supervision to learn representations. This learning paradigm solves the critical problem of utilizing unlabeled graph data. Conventional self-supervised graph learning methods rely on graph data augmentation to generate different views of the input data as self-defined signals. However, the views generated by such an approach contain amounts of identical node features, which leads to the learning of redundant information. To this end, we propose Self-Supervised Graph Learning with Segmented Graph Channels (SGL-SGC) to address the issue. SGL-SGC divides the input graph data across the feature dimensions as Segmented Graph Channels (SGCs). By combining SGCs with data augmentation, SGL-SGC can generate views that vastly reduce the redundant information. We further design a feature-level weight-sensitive loss to jointly accelerate optimization and avoid the model falling into a local optimum. Empirically, the experiments on multiple benchmark datasets demonstrate that SGL-SGC outperforms the state-of-the-art methods in contrastive graph learning tasks. Ablation studies verify the effectiveness and efficiency of different parts of SGL-SGC.