Logistic Regression Confined by Cardinality-Constrained Sample and Feature Selection.

Area of Science:

• Computer Vision
• Machine Learning
• Biomedical Imaging

Background:

• Logistic regression is crucial for vision tasks but struggles with noisy, imbalanced data and irrelevant features.
• Current solutions often use ad-hoc regularization or address only parts of the problem.

Purpose of the Study:

• To develop a mathematically sound logistic regression model for improved feature and sample selection.
• To enhance classification accuracy in vision-based applications despite data challenges.

Main Methods:

• Proposed a logistic regression model incorporating cardinality constraints (l0-norm sparsity) for feature and sample selection.
• Utilized block coordinate descent combined with penalty decomposition to solve non-convex optimization problems.

Main Results:

• The proposed method demonstrated higher accuracy compared to existing techniques on synthetic, image recognition, and neuroimaging datasets.
• Effectively handles noisy data, irrelevant features, and class imbalance by selecting optimal subsets.

Conclusions:

• The novel logistic regression approach offers a robust and accurate solution for challenging vision-based classification tasks.
• Provides a mathematically grounded method for feature and sample selection, outperforming conventional approaches.