Pairwise Constraint-Guided Sparse Learning for Feature Selection.

Area of Science:

• Machine Learning
• Data Mining
• Bioinformatics

Background:

• Feature selection is crucial for creating compact and accurate data representations.
• Lasso and L1-norm regularization are common in supervised feature selection, primarily using class labels.
• Existing methods often overlook pairwise constraints (must-link, cannot-link) which offer valuable, general supervised information.

Purpose of the Study:

• To propose a novel pairwise constraint-guided sparse (CGS) learning method for feature selection.
• To leverage must-link and cannot-link constraints as discriminative regularization terms.
• To develop semi-supervised and ensemble variants of the CGS method.

Main Methods:

• Developed a Constraint-guided Sparse (CGS) learning method incorporating pairwise constraints.
• Introduced semi-supervised CGS utilizing labeled, unlabeled data, and pairwise constraints.
• Created ensemble CGS by combining multiple sets of pairwise constraints.

Main Results:

• The proposed CGS method and its variants demonstrated efficacy in feature selection.
• Experiments were conducted on diverse datasets including UCI, gene expression, neuroimaging, and large-scale attribute data.
• Results showed superior performance compared to established feature selection techniques.

Conclusions:

• Pairwise constraints can be effectively integrated into sparse learning for enhanced feature selection.
• The CGS method and its semi-supervised and ensemble variants offer robust solutions for various classification tasks.
• The proposed methods provide a valuable advancement in leveraging weak supervised information for feature selection.