Refining Latent Homophilic Structures over Heterophilic Graphs for Robust Graph Convolution Networks

Graph convolution networks (GCNs) are extensively utilized in various graphtasks to mine knowledge from spatial data. Our study marks the pioneeringattempt to quantitatively investigate the GCN robustness over omnipresentheterophilic graphs for node classification. We uncover that the predominantvulnerability is caused by the structural out-of-distribution (OOD) issue. Thisfinding motivates us to present a novel method that aims to harden GCNs byautomatically learning Latent Homophilic Structures over heterophilic graphs.We term such a methodology as LHS. To elaborate, our initial step involveslearning a latent structure by employing a novel self-expressive techniquebased on multi-node interactions. Subsequently, the structure is refined usinga pairwisely constrained dual-view contrastive learning approach. Weiteratively perform the above procedure, enabling a GCN model to aggregateinformation in a homophilic way on heterophilic graphs. Armed with such anadaptable structure, we can properly mitigate the structural OOD threats overheterophilic graphs. Experiments on various benchmarks show the effectivenessof the proposed LHS approach for robust GCNs.