Contrastive self-representation learning for data clustering.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Subspace learning is crucial for clustering, with Low-Rank Representation (LRR) showing promise.
• Existing LRR methods overlook data differences and struggle with noise and inter-cluster relationships.

Purpose of the Study:

• To propose a novel Contrastive Self-representation model for Clustering (CSC).
• To address limitations of traditional LRR in handling data dissimilarity and noise.
• To improve the characterization of cluster structures and relationships.

Main Methods:

• Developed a Contrastive Self-representation model (CSC) incorporating similarity and dissimilarity.
• Utilized a specialized loss function to mitigate noise effects.
• Applied an ℓ1,2-norm regularizer for coefficient matrix sparsity and enhanced cluster structure representation.

Main Results:

• The proposed CSC model effectively learns self-representation coefficients.
• CSC demonstrates superior performance in subspace clustering tasks.
• Experiments confirm the model's ability to capture both discriminative information and cluster structures.

Conclusions:

• The Contrastive Self-representation model for Clustering (CSC) offers significant improvements over existing methods.
• CSC provides a robust approach for subspace clustering, adept at handling noise and complex data relationships.
• The method's effectiveness is validated across multiple benchmark datasets.