Prototype based contrastive graph clustering network for reducing false negatives.

Area of Science:

• Graph Machine Learning
• Unsupervised Learning
• Data Mining

Background:

• Contrastive graph clustering methods enhance performance using multi-view augmentation and contrastive loss.
• Self-Supervised Graph Contrastive Learning (SS-GCL) reduces reliance on labeled data but often suffers from false negatives due to pseudo-labeling.

Purpose of the Study:

• To address the limitations of existing SS-GCL methods in graph clustering.
• To propose a novel prototype-driven contrastive graph clustering network that mitigates false negatives and improves clustering effectiveness.

Main Methods:

• A prototype-driven network using data-driven cluster centers (prototypes) to form high-confidence sample sets and aggregate augmented embeddings.
• A cross-view decoupled contrastive learning mechanism employing a mean squared error contrastive loss function solely on positive samples.
• Alignment of augmented positive sample embeddings across views to prevent false negative generation.

Main Results:

• The proposed method effectively prevents the generation of false negative samples.
• Experimental results show superior performance compared to state-of-the-art baseline methods.
• Demonstrated improvements in accuracy and clustering effectiveness across multiple datasets.

Conclusions:

• The prototype-driven contrastive graph clustering network offers a robust solution to the false negative problem in SS-GCL.
• The method achieves state-of-the-art performance, highlighting its potential for advanced graph clustering tasks.