Progressive Self-Supervised Clustering With Novel Category Discovery.

Area of Science:

• Machine Learning
• Pattern Recognition
• Data Analysis

Background:

• Clustering is a fundamental data analysis technique in machine learning and pattern recognition.
• Anchor-based graphs have recently improved graph-based clustering methods.
• Existing methods may require further refinement for optimal clustering performance.

Purpose of the Study:

• To propose a novel progressive self-supervised clustering method with novel category discovery (PSSCNCD).
• To enhance clustering accuracy and performance through a new approach.
• To enable effective data structure analysis in machine learning.

Main Methods:

• A semisupervised framework with novel category discovery guides label propagation.
• A parameter-insensitive anchor-based graph is generated using balanced K-means and hierarchical K-means (BKHK).
• A representative point selection strategy progressively assigns pseudolabels for self-supervised learning.

Main Results:

• The proposed PSSCNCD method demonstrates superior performance compared to existing clustering approaches.
• Experimental results on toy and benchmark datasets validate the effectiveness of the method.
• The approach successfully predicts sample labels for final clustering outcomes.

Conclusions:

• The PSSCNCD method offers a significant advancement in clustering techniques.
• The novel category discovery and progressive pseudolabeling contribute to improved accuracy.
• This method provides a robust solution for complex data clustering challenges.