HGUL Framework for Robust Learning on Heterophilic Heterogeneous Graphs
A new study from arXiv (2604.27387) addresses robust representation learning on heterogeneous graphs with heterophily, where nodes of different types and labels interact non-homophilously. The authors identify structural noise as a critical challenge that degrades model performance. They propose Heterogeneous Graph Unified Learning (HGUL), a unified framework with three modules: a kNN-based graph construction module to recover reliable local neighborhoods, a graph structure learning module to adaptively refine adjacency by filtering noisy edges, and a heterogeneous affinity learning module. The framework jointly handles heterophily and noisy graph structures. The research is published on arXiv under ID 2604.27387.
Key facts
- Heterogeneous graphs with heterophily model complex real-world systems.
- Structural noise significantly degrades model performance.
- HGUL framework proposed to handle heterophily and noisy graph structures.
- Three modules: kNN-based graph construction, graph structure learning, heterogeneous affinity learning.
- kNN module recovers reliable local neighborhoods.
- Graph structure learning module filters noisy edges.
- Published on arXiv with ID 2604.27387.
- Robust representation learning for such graphs was largely unexplored.
Entities
Institutions
- arXiv