Self-Supervised Graph Embedding Clustering.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Manifold learning and K-means clustering are key AI techniques for data analysis.
• Directly combining these models for label learning is a common strategy.
• Existing methods suffer from naive integration, extra hyperparameters, and lack of cluster balance.

Purpose of the Study:

• To develop a meaningful integration of manifold learning and K-means for dimensionality reduction clustering.
• To propose a novel self-supervised framework that unifies these techniques.
• To eliminate the need for additional hyperparameters and ensure cluster balance.

Main Methods:

• A self-supervised manifold clustering framework is proposed, unifying manifold learning and K-means.
• The relationship between K-means and manifold learning is analyzed to build a low-dimensional manifold clustering model.
• The model directly produces a label matrix, which guides manifold structure learning for label-manifold consistency.

Main Results:

• The unified framework achieves dimensionality reduction clustering without extra hyperparameters.
• The ${\ell _{2,p}}$-norm regularization is identified to naturally maintain class balance during clustering, with theoretical proof.
• Experimental results validate the efficiency of the proposed model.

Conclusions:

• The proposed self-supervised framework offers an effective and unified approach to manifold learning and clustering.
• The method addresses limitations of previous integration strategies, providing hyperparameter-free and balanced clustering.
• The findings highlight the utility of ${\ell _{2,p}}$-norm regularization for achieving balanced clustering.