Max-Min Robust Unsupervised Feature Selection via Sparse Subspace.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Feature selection is crucial for reducing data dimensionality and improving model efficiency.
• Existing methods using L2,1-norm regularization face limitations in sparsity and parameter tuning, often leading to suboptimal solutions.
• Unsupervised feature selection is vital for leveraging unlabeled data in complex datasets.

Purpose of the Study:

• To propose a novel max-min robust unsupervised feature selection (MMRUFS) method.
• To address the limitations of existing feature selection algorithms, including sparsity constraints and parameter sensitivity.
• To enhance model robustness and incorporate anomaly detection capabilities within the feature selection process.

Main Methods:

• MMRUFS incorporates both reconstruction and variance terms to preserve data information and enhance dispersion.
• Utilizes L2,0-norm constraint on the transformation matrix for direct optimal feature subset selection, avoiding parameter tuning.
• Employs a designed mark weight vector for robust handling of normal samples and outliers, enabling anomaly detection.
• Guarantees convergence through a surrogate matrix-based solution approach.

Main Results:

• MMRUFS effectively retains original data information while increasing data dispersion.
• The L2,0-norm constraint facilitates direct selection of optimal feature subsets, simplifying the process.
• The method demonstrates robustness by differentiating between normal samples and outliers, aiding anomaly detection.
• Experimental results confirm that MMRUFS outperforms existing feature selection algorithms on various real-world datasets.

Conclusions:

• MMRUFS offers a robust and efficient unsupervised feature selection approach.
• The method's ability to handle outliers and avoid parameter tuning makes it a practical solution.
• MMRUFS demonstrates superior performance compared to traditional feature selection techniques, highlighting its potential for diverse applications.