Efficient cross-validation traversals in feature subset selection.

Area of Science:

• Machine Learning
• Computational Biology
• Bioinformatics

Background:

• Sparse and robust classification models identify predictive patterns and generate hypotheses.
• Feature selection is crucial but computationally challenging due to large search spaces.
• Current methods have limited coverage, even for low-dimensional data.

Purpose of the Study:

• To present methods for reducing the computational complexity of feature selection criteria.
• To enhance the efficiency and coverage of feature screening in classification.
• To enable higher-dimensional analyses by reducing preparation costs.

Main Methods:

• Developed methods to reduce computational complexity of feature selection criteria.
• Integrated a parallelizable cross-validation traversal strategy with distance-based classifiers.
• Methods are compatible with any product distance or kernel.

Main Results:

• Achieved significant reduction in computational complexity for high-dimensional subsets.
• Demonstrated approximately a 15-fold increase in generating distance matrices for feature combinations.
• Evaluated performance, runtime, and fitness landscape on public datasets.

Conclusions:

• The proposed methods significantly improve the efficiency and coverage of feature selection.
• Enables more comprehensive evaluations, even in low-dimensional settings.
• Advances the potential of sparse classification models for pattern discovery and hypothesis generation.