Benchmarking relief-based feature selection methods for bioinformatics data mining | JoVE Visualize

Area of Science:

Biomedical data mining
Bioinformatics
Machine learning

Background:

Modern biomedical data mining demands feature selection methods capable of handling large-scale 'omics' data, noisy conditions, and complex associations like gene-gene interactions.
Existing methods often struggle with the adaptability required for diverse data types (genetic variants, gene expression, clinical data) and computational tractability.

Purpose of the Study:

To examine and expand Relief-Based Algorithms (RBAs) for robust feature selection in biomedical data mining.
To introduce and evaluate a novel RBA, MultiSURF, within the open-source ReBATE framework.
To compare the performance of RBAs against other established feature selection methods using a comprehensive genetic simulation study.

Main Methods:

Implementation and expansion of Relief-Based Algorithms (RBAs) within the open-source ReBATE (Relief-Based Algorithm Training Environment) framework.
A comprehensive genetic simulation study comparing existing RBAs, the proposed MultiSURF algorithm, and other established feature selection methods.
Evaluation across diverse problem types, including classification vs. regression, discrete vs. continuous features, missing data, multiple classes, and class imbalance.

Main Results:

RBAs demonstrate flexibility, efficiency, and power in differentiating relevant features with univariate, multivariate, epistatic, or heterogeneous associations.
Expansions of RBAs confirm efficacy for various data characteristics and problem types, including handling missing data and class imbalance.
The study identifies limitations of specific RBAs and highlights MultiSURF as yielding the most reliable feature selection performance across a wide range of problem types, with MultiSURF* excelling at identifying 2-way interactions.

Conclusions:

Relief-Based Algorithms are highly effective and adaptable feature selection tools for modern biomedical data mining.
The ReBATE framework and the MultiSURF algorithm provide valuable advancements for analyzing complex biological datasets.
The findings offer guidance for selecting appropriate feature selection methods based on specific data characteristics and research objectives.