Abstract
Graphs provide a fundamental way to model relationships between entities and are central to numerous machine learning tasks. Standard graph-based methods often assume the provided graph structure is both accurate and complete. However, real-world graphs frequently suffer from noise and sparsity, negatively impacting downstream tasks like node classification and clustering. While graph representation learning has advanced significantly, many methods implicitly assume graph homophily (connections predominantly between nodes of the same class), struggling when faced with heterophily (connections predominantly between different classes). This paper introduces a novel method, Resilient Graph Learning for Heterophily (RGLH), designed to learn high-quality graph structures directly from potentially heterophilic data. RGLH first employs a high-pass graph filter to enhance node distinctiveness relative to neighbors, encoding structural patterns into features suitable for heterophilic scenarios. Subsequently, it learns a robust graph using an adaptive norm capable of handling varying noise levels and incorporates a novel regularizer inspired by contrastive learning to refine the graph structure without requiring data augmentation. Extensive experiments on clustering and semi-supervised classification tasks using benchmark heterophilic datasets demonstrate the effectiveness of RGLH. Notably, our relatively simple approach achieves superior performance compared to several complex deep learning methods, particularly in challenging heterophilic environments.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
"Learning Resilient Graph Structures in Heterophilic Settings", Technical Disclosure Commons, (May 04, 2025)
https://www.tdcommons.org/dpubs_series/8078