Information-Enhanced Graph Neural Network for Transcending Homophily Barriers

Homophily and heterophily are intrinsic properties of graphs that describe whether linked nodes share similar properties. While Message Passing Neural Networks (MPNNs) have shown remarkable success in node classification tasks, their performance often deteriorates within specific homophily ranges, which we term the gray area. In this work, we identify and theoretically demonstrate the challenges faced by MPNNs in this gray area, highlighting the limitations of existing approaches in addressing it. To overcome these limitations, we propose the INformation-enhanced Graph Neural Network (INGNN), which introduces a novel framework that integrates three complementary features-ego-node features, graph structure features, and aggregated neighborhood features-through an adaptive feature fusion mechanism based on bi-level optimization. This design enables INGNN to transcend the homophily barriers and generalize effectively across the entire homophily spectrum. We validate the effectiveness of INGNN through extensive experiments on both synthetic and real-world datasets with different graph homophily. Specifically, INGNN outperforms 12 state-of-the-art MPNNs with an average rank of 1.78 on 9 real node classification datasets. Our ablation studies further show experimental evidence of how the integrated features contribute to the model’s performance under different homophily settings. INGNN is open-sourced and available at https://github.com/xl1990/ingnn.