Heterophily-Aware Representation Learning on Heterogeneous Graphs.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Mining

Background:

• Real-world graphs are complex, featuring global heterogeneity and local heterophily.
• Existing graph neural networks (GNNs) struggle with heterophily, especially in heterogeneous graphs.
• Heterogeneous graph neural networks (HGNNs) often inherit limitations from GNNs, failing to generalize to heterophilic graphs.

Purpose of the Study:

• To investigate heterophily in heterogeneous graphs.
• To propose metrics for quantifying heterophily in heterogeneous graphs.
• To develop a heterophily-aware HGNN capable of handling both homophilic and heterophilic heterogeneous graphs.

Main Methods:

• Identified heterophily in heterogeneous graphs using metapaths.
• Proposed two quantitative metrics for heterophily levels.
• Developed Hetero$^{2}$2Net, incorporating masked metapath and label prediction tasks.
• Evaluated on five real-world heterogeneous graph benchmarks.

Main Results:

• Empirical investigations revealed HGNNs struggle with heterophilic heterogeneous graphs.
• Hetero$^{2}$2Net demonstrated superior performance in semi-supervised node classification.
• The model successfully scaled to a large industrial-scale graph (13M nodes, 157M edges).

Conclusions:

• Heterophily is a critical challenge in heterogeneous graph analysis.
• Hetero$^{2}$2Net effectively addresses heterophily in HGNNs.
• The proposed model offers a scalable and effective solution for complex real-world heterogeneous graphs.