Hellinger distance-based stable sparse feature selection for high-dimensional class-imbalanced data.

Area of Science:

• Machine Learning
• Bioinformatics
• Data Science

Background:

• Class-imbalance learning is crucial for high-dimensional data analysis in various scientific fields.
• Traditional feature selection methods struggle with imbalanced datasets, necessitating new approaches.
• Effective feature selection improves classification performance and biomarker discovery in complex data.

Purpose of the Study:

• To develop a stable and sparse feature selection algorithm for high-dimensional class-imbalanced data.
• To address the limitations of existing feature selection techniques in handling imbalanced datasets.
• To provide an efficient method for identifying key features in complex biological and scientific data.

Main Methods:

• Proposed the sssHD algorithm, combining Hellinger distance (HD) with sparse regularization.
• Utilized HD's class-insensitive and translation-invariant properties for robust feature selection.
• Evaluated sssHD on simulated data and five gene expression datasets, comparing it with existing methods.

Main Results:

• The HD-based selection algorithm effectively identifies key features and controls false discoveries in imbalanced learning.
• sssHD demonstrated highly competitive performance across five assessment metrics compared to existing procedures.
• The algorithm showed minimal performance differences with or without re-balance preprocessing.

Conclusions:

• sssHD is a practical, simple, and generalizable feature selection method for high-dimensional imbalanced data.
• The algorithm's flexibility allows extension with different preprocessing, regularization, and classifiers.
• Offers a valuable alternative for feature selection in diverse class-imbalanced learning scenarios.