When Heterophily Meets Heterogeneous Graphs: Latent Graphs Guided Unsupervised Representation Learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Unsupervised heterogeneous graph representation learning (UHGRL) is crucial for unlabeled real-world graphs.
• Heterophily, where connected nodes are dissimilar, is prevalent but often overlooked in UHGRL.
• Existing methods struggle to capture the nuances of heterophily in heterogeneous graph data.

Purpose of the Study:

• To address the challenge of semantic heterophily in unsupervised heterogeneous graph representation learning.
• To propose a novel framework, LatGRL, capable of handling both homophily and heterophily.
• To develop a scalable solution for large-scale real-world graph datasets.

Main Methods:

• Introduced a similarity mining method coupling global structures and attributes.
• Constructed fine-grained homophilic and heterophilic latent graphs (LGs) to guide representation learning.
• Developed an adaptive dual-frequency semantic fusion mechanism for node-level heterophily and a scalable implementation.

Main Results:

• LatGRL effectively handles semantic heterophily in unsupervised heterogeneous graph representation learning.
• The framework demonstrates superior performance on benchmark datasets.
• The proposed methods validate the effectiveness and efficiency of the LatGRL framework.

Conclusions:

• LatGRL provides a robust solution for unsupervised heterogeneous graph representation learning in the presence of heterophily.
• The framework's ability to mine and utilize latent graphs significantly enhances representation quality.
• The scalable implementation makes LatGRL applicable to large-scale, real-world graph data.