Investigation on Trustworthy Graph-Based Learning Methods Beyond Homophily Receives RGC FDS Funding

We are delighted to announce that our latest project, “Towards Graph-Based Learning System Under Diverse Graph Homophily: Method, Vulnerability and Robustness” (UGC/FDS13/E06/25), has been funded by the Research Grants Council (RGC) under the Faculty Development Scheme (FDS) for the 2025/26 academic year.

“Investigating the Trustworthy Graph-Based Learning Methods Under Diverse Graph Homophily”, Dr. Zhu, Yulin (PI), Assistant Professor Department of Computer Science, Prof. Lo, Wai Lun (CoI), Professor Department of Computer Science, Prof. Fan, Xiaodan (CoI), Professor, Dept. of Statistics, The Chinese University of Hong Kong, $400,300 (24 months), UGC/FDS13/E06/25.

Abstract

Graph-based learning methods are powerful AI tools that excel at learning from relational data, like social networks, recommendation systems, transportation networks, and transaction networks. However, they face significant trust issues because they are vulnerable to malicious adversarial noises and often get confused by task-irrelevant patterns, limiting their applications in the complex real-world scenarios. While previous research has improved trustworthiness by enhancing generalization and robustness, these efforts primarily relied on the “homophily” assumption—where similar nodes are tended to be connected—which works well for some graphs but fails dramatically for others where connected nodes are distinct (heterophilic graphs), narrowing their practical application. To address this problem, our preliminary studies have uncovered a key insight: the similarity between aggregated node embeddings can prominently distinguish reliable connections from task-irrelevant ones, providing a new foundation for graph data augmentation that works beyond homophily. Building on this, we aim to develop a more universal and automated trustworthy graph learning system that can handle various challenges—including noisy data, different learning settings, and diverse graph types—to ultimately deliver robust and reliable predictions across real-world scenarios.

Principle Investigator

Hong Kong Chu Hai College