An Unsupervised Feature Extraction Approach Based on Self-Expression.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Current feature extraction methods often lack interpretability in projection learning.
• Obtained low-dimensional features may not be as explanatory as original sample features.

Purpose of the Study:

• To propose an unsupervised feature extraction algorithm with enhanced interpretability.
• To integrate feature extraction and selection into a single framework.

Main Methods:

• Developed block diagonal projection (BDP) based on self-expression.
• Imposed L2,1 norm constraint on the projection matrix for row sparsity.
• Introduced a block diagonal representation regular term.

Main Results:

• BDP enhances projection matrix interpretability by selecting discriminative features and suppressing irrelevant ones.
• The algorithm effectively excavates correlations between samples and features.
• Demonstrated effectiveness in linear dimensionality reduction and classification tasks.

Conclusions:

• BDP offers a superior approach to traditional feature extraction methods.
• The algorithm improves accuracy in pattern recognition tasks like clustering and classification.
• BDP provides a more interpretable and effective solution for unsupervised feature extraction.