Abstract
Graph Neural Networks (GNNs) have demonstrated significant potential for learning from graphstructured data. However, their performance often degrades on heterophilous graphs, where linked nodes exhibit dissimilar features. In this paper, we introduce the Adaptive Relational Graph Network (ARGN), a novel GNN architecture designed specifically to handle heterophily. ARGN utilizes a relational attention mechanism to capture complex relationships between nodes, coupled with an adaptive neighborhood selection strategy to dynamically identify relevant neighbors. Multi-hop information propagation is then used to enhance the node representations. We provide a detailed algorithm and implementation of ARGN using PyTorch. Our experiments on several benchmark heterophilous datasets demonstrate that ARGN significantly outperforms state-of-the-art GNN models. We conclude with a discussion of these results and suggestions for future research.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
"Adaptive Relational Graph Network", Technical Disclosure Commons, (February 03, 2025)
https://www.tdcommons.org/dpubs_series/7800