Deep Graph Clustering Framework Based on Confidence-Guided Graph Enhancement and Dual-Negative Sample Contrastive...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Mining

Background:

• Attributed graph clustering is unsupervised node partitioning using graph topology and attributes.
• Current deep methods struggle with local bias, noisy graph editing, and distinguishing hard samples.

Purpose of the Study:

• To propose a deep graph clustering framework, CGEN, addressing limitations of existing methods.
• To enhance clustering performance by integrating graph enhancement and contrastive learning.

Main Methods:

• Developed a local-global dual-view representation learning module to fuse neighborhood attributes and global topology.
• Implemented a confidence-guided conservative graph editing mechanism with multiple constraints and progressive updates.
• Employed a dual-negative sample contrastive learning strategy to improve discriminative ability for hard samples.

Main Results:

• CGEN demonstrated highly competitive performance on four benchmark datasets.
• The framework outperformed state-of-the-art methods across core clustering metrics.
• Effectiveness was validated in overcoming local structural bias, graph editing noise, and hard sample limitations.

Conclusions:

• CGEN offers a robust solution for attributed graph clustering.
• The proposed methods effectively enhance graph representation and discriminative power.
• CGEN significantly advances the state-of-the-art in unsupervised graph clustering.