A Variance Minimization Criterion to Feature Selection Using Laplacian Regularization.

Area of Science:

• Machine Learning
• Data Science
• Computer Science

Background:

• High-dimensional data presents challenges in information processing tasks.
• Unsupervised learning lacks class labels, complicating feature selection.
• Existing methods struggle with effective feature identification in label-free scenarios.

Purpose of the Study:

• To propose novel feature selection algorithms for unsupervised learning.
• To address the difficulty of feature selection without class labels.
• To enhance clustering, classification, and retrieval in high-dimensional datasets.

Main Methods:

• Developed algorithms based on Laplacian regularized least squares.
• Minimized expected prediction error by optimizing the parameter covariance matrix.
• Utilized trace and determinant operators, inspired by experimental design.

Main Results:

• Proposed algorithms effectively select meaningful feature subsets.
• Demonstrated superiority over existing methods through extensive experiments.
• Achieved efficient computation for the optimization problems.

Conclusions:

• The novel feature selection algorithms are effective for unsupervised learning.
• The proposed methods offer a significant improvement for high-dimensional data analysis.
• These algorithms facilitate better performance in downstream tasks like classification and retrieval.