Subspace Clustering via Structured Sparse Relation Representation.

Area of Science:

• Data Science
• Machine Learning
• Computer Vision

Background:

• Classical spectral clustering struggles with noisy real-world data, failing to accurately reveal subspace structures.
• Existing affinity graph construction methods in subspace clustering are sensitive to data corruptions.

Purpose of the Study:

• To propose a novel approach for subspace clustering by addressing limitations in affinity graph construction.
• To introduce the concept of 'relation reconstruction' for more robust data representation.

Main Methods:

• Developed 'relation reconstruction' where neighborhood relations represent data sample membership.
• Proposed Sparse Relation Representation (SRR) and Structured Sparse Relation Representation (SSRR) methods.
• Presented an optimization algorithm for SRR and SSRR, analyzing computational complexity and convergence.

Main Results:

• SRR and SSRR methods demonstrated superior performance in subspace clustering tasks.
• Experiments on diverse datasets confirmed the effectiveness of the proposed approaches.
• The neighborhood relation representation proved robust against data noise and corruption.

Conclusions:

• Relation reconstruction offers a more faithful way to construct affinity graphs for subspace clustering.
• SRR and SSRR provide significant improvements over traditional spectral clustering methods, especially on noisy data.
• The proposed methods enhance the discovery of intrinsic subspace structures in complex datasets.