Robust Graph Neural Networks via Adaptive Aggregator Selection and Long-Range Dependency Modeling

Deep learning on graphs has received increasing attention in recent years, and Graph Neural Networks (GNNs) have achieved remarkable performance across various tasks on graphs like node classification and link prediction, etc. However, recent studies show that GNNs can be vulnerable to adversarial graph structure perturbations. To address this issue, we propose GalNN, a GNN architecture based on adaptive aggregator selection and long-range dependency modeling, which is robust to poisoning attack on both homophily graphs and heterophily graphs. The general principle of adaptive aggregator selection is to learn each neighborhood’s profile via the centering node’s degree and neighborhood feature variance to guide the aggregation with a set of aggregators. Long-range dependency modeling adopts the manifold learning idea to locate similar nodes in the neighboring regions of a 1D manifold and simulates the aggregation operation with 1D convolution. Our results show that the learned neighborhood profiles can differentiate perturbed neighborhoods from clean neighborhoods on attacked homophilic graphs, and the long-range dependency modeling can make the original heterophily graphs more homophilic. Fusing the information from adaptive aggregator selection and long-range dependency modeling realizes robust node embedding learning.