What Contributes More to the Robustness of Heterophilic Graph Neural Networks?

In recent years, graph neural networks (GNNs) have gained significant attention due to their outstanding performance on graph-related tasks by utilizing neighborhood aggregation. However, traditional GNNs are primarily designed based on the homophily assumption, which means that they show poor performance on heterophilic networks where dissimilar nodes prefer to connect. To address this issue, several heterophilic GNNs have been proposed that employ techniques, such as extending local neighbors and improving GNN architectures. From another perspective, recent studies have shown that, unlike homophilic GNNs, heterophilic GNNs exhibit higher robustness against graph adversarial attacks. In these attacks, the attackers try to inject small perturbations into the graph. Therefore, in this study, we delve into the core designs of heterophilic GNNs, including high-order neighbor, potential neighbor, ego-neighbor separation, and interlayer combination designs. We further analyze the influence of these key designs on the robustness of GNNs. We conducted comprehensive experiments to compare the impact of different designs on baseline and real-world GNN models. The findings of this work can serve as a reference for future studies aiming to design more robust and universal GNNs.