Graph Clustering with High-Order Contrastive Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Theory

Background:

• Graph clustering is a key unsupervised learning task.
• Contrastive learning has advanced graph clustering but faces challenges.
• Existing methods suffer from manual augmentations causing semantic drift and feature-level focus neglecting graph structure.

Purpose of the Study:

• To propose a novel method, Graph Clustering with High-Order Contrastive Learning (GCHCL), to overcome limitations in current graph clustering techniques.
• To enhance unsupervised graph clustering by integrating higher-order structural information and automatic view generation.

Main Methods:

• GCHCL constructs two views using Laplacian smoothing with varied normalizations and employs a structure alignment loss.
• It builds a contrastive similarity matrix from structure-based similarities, aligning it with an identity matrix for enhanced neighborhood learning.
• The method learns clustering-friendly embeddings directly, eliminating the need for a separate clustering module and enabling scalability.

Main Results:

• GCHCL demonstrated significant effectiveness across five datasets.
• The model achieved an average accuracy improvement of 3% over strong baselines on small and medium datasets.
• On the largest dataset, GCHCL attained 81.92% accuracy, overcoming out-of-memory issues faced by other methods.

Conclusions:

• GCHCL offers a robust and scalable solution for graph clustering by leveraging high-order structural information.
• The proposed method effectively addresses semantic drift and improves performance by integrating structure-level contrastive learning.
• GCHCL provides superior accuracy and efficiency, particularly for large-scale graph clustering tasks.