Researchers announced on Apr. 13 a new training technique that improves the accuracy of graph neural networks (GNNs), which are artificial intelligence systems used in fields such as drug discovery and weather forecasting.
GNNs process data structured as graphs, where nodes represent data points and edges indicate relationships between them. These relationships can connect similar nodes, known as homophily, or dissimilar ones, referred to as heterophily. The ability of GNNs to model complex connections makes them valuable for applications ranging from social networks to molecular structures.
Training GNNs is challenging, especially when using self-supervised learning where no node labels are provided. “If none of the nodes are labeled, the GNN can see that there are edges between nodes but has trouble distinguishing between homophilic edges and heterophilic edges,” said Tianfu Wu, senior author of the study and associate professor of electrical and computer engineering at North Carolina State University. “And this problem is especially pronounced in heterophilic graphs, meaning graphs that have more heterophilic relationships than homophilic ones. That’s the problem we’re addressing with this work.”
The researchers introduced a framework called HarmonyGNN that enhances GNN performance on both homophilic and heterophilic graphs without compromising accuracy on either type. When tested against 11 benchmark datasets, HarmonyGNN matched state-of-the-art results for seven homophilic graphs and set new records for four heterophilic ones, with improvements ranging from 1.27% to 9.6%. “This is a significant advance for GNN training,” Wu said. “In addition, the HarmonyGNN framework also improved the computational efficiency of the training.” Relevant code has been made available on GitHub.
The findings will be presented at the Fourteenth International Conference on Learning Representations (ICLR2026) scheduled for April 23-27 in Rio de Janeiro, Brazil. The paper’s first author is Rui Xue, a Ph.D. student at NC State University.
Support for this research came from grants provided by the Army Research Office and National Science Foundation.
