Structure-preserving contrastive graph clustering with dual-channel label alignment.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Mining

Background:

• Contrastive Graph Clustering (CGC) has rapidly advanced, but current methods face challenges with data augmentation altering semantics and overlooking discriminative unsupervised information.
• Existing approaches often neglect dynamic structural relationships by relying on static neighborhood connections, impacting clustering performance.

Purpose of the Study:

• To propose a Structure-preserving Contrastive Graph Clustering approach with Dual-channel Label Alignment (SCGC-DLA).
• To address limitations in view generation, sample pairing, and dynamic structural relationship learning in CGC.

Main Methods:

• Utilizes low-pass and hybrid graph filters for reliable, complementary data augmentation views.
• Constructs a structure-preserving matrix using edge betweenness centrality (EBC) to capture topological relationships.
• Employs non-dominated sorting theory for dual-channel clustering distribution to generate high-confidence pseudo-labels aligned with latent semantic labels.
• Applies a self-supervised learning scheme to guide the overall network for final clustering.

Main Results:

• The proposed SCGC-DLA method demonstrates robustness and effectiveness across five benchmark datasets.
• Achieves superior clustering performance compared to several state-of-the-art methods.
• Validates the benefits of structure preservation and dual-channel label alignment in CGC.

Conclusions:

• SCGC-DLA effectively overcomes key challenges in contrastive graph clustering.
• The method's novel approach to augmentation, structure preservation, and label alignment leads to improved clustering results.
• This work offers a significant advancement in the field of graph clustering.