An End-to-End Deep Graph Clustering via Online Mutual Learning

Area of Science:

• Machine Learning
• Data Mining
• Graph Theory

Background:

• Deep graph models typically use separate stages for neural network optimization and data aggregation.
• Existing methods face limitations where clustering results cannot guide neural network training.
• Complex matrix computations in aggregation lead to high computational burdens.

Purpose of the Study:

• To propose a unified deep graph clustering (UDGC) model for end-to-end optimization.
• To address the limitations of separate optimization stages in deep graph clustering.
• To reduce computational complexity in deep graph clustering tasks.

Main Methods:

• Developed a unified deep graph clustering (UDGC) model utilizing online mutual learning.
• Extracted deep graph representations and node topological knowledge in a deep embedding subspace.
• Employed a local preserving loss for embedding aggregation and clustering assignment generation.
• Trained a neural layer to fit clustering results and optimized the model end-to-end.

Main Results:

• The UDGC model achieves end-to-end clustering assignment generation.
• Significant reduction in computational complexity compared to traditional methods.
• Demonstrated superior performance through extensive experiments.

Conclusions:

• The proposed UDGC model effectively unifies deep embedding extraction and clustering.
• Online mutual learning enables efficient end-to-end optimization of deep graph clustering.
• The UDGC model presents a superior and computationally efficient approach for graph clustering tasks.