Low-rank and sparse embedding for dimensionality reduction.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Subspace learning (SL) is crucial for dimensionality reduction.
• Existing SL methods often lack robustness and generalization capabilities.
• Noise and data corruption can significantly degrade SL performance.

Purpose of the Study:

• To propose a robust subspace learning framework for dimensionality reduction.
• To extend existing SL methods with low-rank and sparse embedding (LRSE).
• To enhance overall optimum, robustness, and generalization of SL.

Main Methods:

• Developed a low-rank and sparse embedding (LRSE) framework.
• Jointly performed reconstruction coefficient matrix learning and SL.
• Utilized a sparse matrix to model and mitigate noise.

Main Results:

• LRSE captures both global subspaces and local geometric structures.
• Achieved superior performance over conventional SL methods in unsupervised and supervised scenarios.
• Demonstrated considerable improvement in classification accuracy, particularly in unsupervised tasks.
• Experiments validated the superiority of derived SL methods on various datasets, including corrupted data.

Conclusions:

• The proposed LRSE framework offers enhanced robustness and generalization for dimensionality reduction.
• LRSE provides a unified framework from which seven specific SL methods can be derived.
• The study offers new insights into subspace learning through experimental analysis.