Unsupervised feature selection based on variance-covariance subspace distance.

Area of Science:

• Machine Learning
• Data Science
• Statistics

Background:

• Subspace distance is crucial for feature selection, identifying representative feature subsets.
• Existing methods often neglect intrinsic data statistics, limiting their effectiveness.
• A gap exists in feature selection methods that leverage both subspace properties and data correlations.

Purpose of the Study:

• To propose a novel unsupervised feature selection framework utilizing Variance-Covariance subspace distance.
• To address the limitations of existing subspace distance methods by incorporating feature correlations.
• To perform dimensionality reduction and subspace learning simultaneously.

Main Methods:

• Introduced "Variance-Covariance subspace distance" to leverage feature correlations.
• Developed a framework to identify feature subsets with minimum norm Variance-Covariance matrices.
• Provided an efficient update algorithm with convergence analysis for optimization.

Main Results:

• The proposed method effectively handles dimensionality reduction and subspace learning.
• It successfully excludes features with minimal variance.
• Experiments on nine benchmark datasets showed superior performance compared to state-of-the-art methods.

Conclusions:

• The Variance-Covariance subspace distance offers a powerful approach for unsupervised feature selection.
• The framework enhances feature selection by considering feature correlations and intrinsic data variance.
• This method presents a significant advancement in unsupervised learning and dimensionality reduction.