Fast self-supervised discrete graph clustering with ensemble local cluster constraints.

Area of Science:

• Data mining
• Image processing
• Machine learning

Background:

• Spectral clustering (SC) is a widely used graph-based clustering algorithm.
• Traditional SC methods often neglect prior information, hindering accuracy in unsupervised settings.
• Existing algorithms require extensive hyperparameter tuning and full graph construction, increasing computational costs.

Purpose of the Study:

• To propose a simple, fast, self-supervised discrete graph clustering (FSDGC) algorithm.
• To address limitations of traditional spectral clustering, including hyperparameter tuning and computational expense.
• To enhance clustering performance by incorporating prior information effectively.

Main Methods:

• Introduced novel self-supervised information based on ensemble local cluster constraints.
• Utilized an anchor graph technique for efficient large-scale dataset processing.
• Employed a fast coordinate ascent (CA) optimization method for discrete indicator matrices.

Main Results:

• The FSDGC method demonstrated efficient and effective clustering performance.
• Self-supervised constraints improved the accuracy of clustering results.
• The anchor graph technique enabled scalability to large datasets.

Conclusions:

• FSDGC offers an efficient and accurate alternative to traditional spectral clustering methods.
• The integration of self-supervised learning and anchor graphs enhances clustering in data mining and image processing.
• The proposed method reduces computational overhead and simplifies the clustering process.