Robust Joint Graph Sparse Coding for Unsupervised Spectral Feature Selection.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Feature selection is crucial for dimensionality reduction and improving model performance.
• Preserving local data structures is essential for effective analysis.
• Existing methods may be sensitive to outliers and noise.

Purpose of the Study:

• To propose a novel unsupervised spectral feature selection model.
• To embed a graph regularizer within joint sparse regression.
• To enhance data representation by preserving local structures.

Main Methods:

• Developed a joint graph sparse coding (JGSC) model.
• Incorporated subspace learning and joint sparse regression.
• Extended JGSC to a robust version (RJGSC) using a robust loss function.
• Designed and proved the convergence of a new optimization solution.

Main Results:

• JGSC and RJGSC effectively preserve local data structures.
• RJGSC demonstrates robustness against outliers.
• Both models significantly outperform state-of-the-art algorithms in k-nearest neighbor classification.

Conclusions:

• The proposed JGSC and RJGSC models offer superior performance in unsupervised spectral feature selection.
• These methods provide effective solutions for handling local data structures and outliers.
• The models show promise for various machine learning applications requiring robust feature selection.