Inventor(s)

Abstract

Graph Neural Networks (GNNs) have shown impressive results in analyzing graph-structured data. However, their performance is often compromised when applied to heterophilous graphs, where connected nodes tend to have dissimilar features. This paper introduces the Structure-Aware Graph Neural Network (SAGNN), a novel GNN architecture designed to handle heterophilous graphs by explicitly encoding and leveraging structural information. SAGNN uses a structure encoding module to generate structural node features, followed by an adaptive message aggregation mechanism based on structural similarity, and finally a structure-aware combination module to generate final embeddings. We provide the algorithm of SAGNN and also a Python implementation with PyTorch. Extensive experimental results on benchmark heterophilous graph datasets demonstrate that SAGNN significantly outperforms both standard GNN models and existing state-of-the-art heterophilyfocused methods.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Share

COinS