Xin Zheng, Yi Wang, Yixin Liu, Ming Li, Miao Zhang, Di Jin, Philip S. Yu, Shirui Pan
A comprehensive survey that systematically categorizes and analyzes graph neural network (GNN) models specialized for graphs with heterophily.
Most GNNs rely on the homophily assumption that nodes of the same class are more likely to be connected. However, in real-world graphs, heterophily (nodes with different labels tend to be linked) is common, significantly degrading the performance of existing GNNs. Thus, a systematic understanding and classification of GNN models that can effectively learn on heterophilic graphs is needed.
This paper organizes existing heterophilic GNN models into a systematic taxonomy. It analyzes the design principles, strengths, weaknesses, and applicable graph characteristics of each model. Additionally, it discusses the relationship between heterophily and various graph research domains (e.g., graph classification, link prediction, anomaly detection) and suggests future research directions.
As the first comprehensive survey in the field of heterophilic graph learning, it provides a useful benchmark for researchers to understand existing methods and develop new ones. It also connects heterophily with diverse graph research domains, promoting interdisciplinary progress.