Similarity preserving low-rank representation for enhanced data representation and effective subspace learning

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Latent Low-Rank Representation (LatLRR) is effective for subspace recovery and feature extraction.
• LatLRR struggles to preserve the locality of similar principal and salient features during optimization.
• This limitation hinders optimal performance in tasks requiring feature similarity preservation.

Purpose of the Study:

• To introduce Regularized Low-Rank Representation (rLRR) to overcome LatLRR's limitations.
• To enhance feature representation by preserving local feature similarities.
• To develop an unsupervised subspace learning framework, Low-rank Similarity Preserving Projections (LSPP), based on rLRR.

Main Methods:

• Proposed Regularized Low-Rank Representation (rLRR) by incorporating Laplacian regularization into LatLRR.
• rLRR decomposes data matrices into low-rank matrices while preserving feature similarities.
• Developed Low-rank Similarity Preserving Projections (LSPP) for unsupervised feature learning using rLRR outputs, with a supervised extension discussed.

Main Results:

• rLRR effectively preserves similarities of principal and salient features, grouping correlated features.
• The proposed rLRR and LSPP frameworks demonstrate enhanced robustness in representation and decomposition.
• Experimental results on real images show the superiority of rLRR and LSPP over existing state-of-the-art algorithms.

Conclusions:

• rLRR significantly improves upon LatLRR by preserving local feature similarities.
• LSPP provides an effective unsupervised and supervised framework for discriminant subspace learning.
• The proposed methods offer superior performance in robust representation and feature extraction tasks.