Robust Subcluster Search and Mergence Clustering.

Area of Science:

• Data Science
• Machine Learning
• Artificial Intelligence

Background:

• Graph-based clustering segments data by partitioning similarity graphs.
• Existing methods struggle with real-world data's inherent noise and outliers.
• Current approaches often derive clusters directly from learned graphs, requiring strict internal data distribution.

Purpose of the Study:

• Introduce a novel clustering model, Robust Subcluster Search and Mergence (RSSM).
• Address limitations of existing graph-based clustering methods, particularly regarding outlier handling.
• Improve the quality of the learned graph for more effective clustering.

Main Methods:

• RSSM leverages outliers, inspired by positive-incentive noise (Pi-Noise), for structure learning.
• It identifies subcentroids by searching an imbalanced residue distribution, separating inliers from outliers.
• A subcluster similarity graph is constructed to guide the merging of identified subclusters.

Main Results:

• Subclusters identified by subcentroids exhibit tighter connections among normal samples.
• RSSM effectively utilizes outliers to refine the graph structure for clustering.
• Experimental results validate the rationality and superiority of the RSSM model.

Conclusions:

• RSSM offers a robust approach to graph-based clustering by effectively handling outliers.
• The simultaneous search and mergence of subclusters, aided by outliers, leads to improved clustering performance.
• The proposed method generates a more suitable graph representation for clustering tasks.