Self-Supervised Graph Representation Learning via Information Bottleneck

Graph representation learning has become a mainstream method for processing network structured data, and most graph methods rely heavily on labeling information downstream tasks. Since labeled is rare in the real world, adopting self-supervised to solve neural problem significant challenge. Currently, existing approaches attempt maximize mutual learning, which leads large amount of redundant thus affects performance Therefore, bottleneck (SGIB) proposed this paper uses symmetry asymmetry graphs establish comparative introduces theory as loss training model. This model extracts common features both views independent each view by maximizing estimation between local high-level one global summary vector other view. It also removes not relevant target task minimizing representations two views. Based extensive experimental results three public datasets large-scale datasets, it been shown that SGIB can learn higher quality node several classical analysis experiments such classification clustering be improved compared models an unsupervised environment. In addition, in-depth experiment designed analysis, show alleviate over-smoothing certain extent. we infer from different would effective improvement tasks through introducing remove information.