A Multicriteria Approach to Find Predictive and Sparse Models with Stable Feature Selection for High-Dimensional Data

Area of Science:

• Bioinformatics
• Computational Biology
• Statistical Genetics

Background:

• High-dimensional data, particularly genetic data, presents challenges for predictive modeling.
• Model interpretability and feature selection stability are crucial for biological insights in bioinformatics.

Purpose of the Study:

• To evaluate and identify optimal criteria for fitting predictive models to high-dimensional datasets.
• To compare various feature selection stability measures and assess their suitability for genetic data analysis.
• To investigate the trade-offs between classification accuracy, feature selection stability, and model complexity.

Main Methods:

• Comparison of multiple feature selection stability measures.
• Assessment of stability measures based on theoretical and empirical properties.
• Utilizing Pareto fronts to analyze multi-objective optimization (accuracy, stability, feature count).
• Focus on measures incorporating corrections for chance and the number of selected features.

Main Results:

• Pearson correlation demonstrated superior theoretical and empirical properties as a stability measure.
• Stability assessment requires measures with corrections for chance or the number of features selected.
• Pareto front analysis revealed that models with stable, reduced feature sets can achieve high predictive accuracy.

Conclusions:

• Pearson correlation is recommended for assessing feature selection stability in high-dimensional genetic data.
• Effective predictive models in bioinformatics must prioritize accuracy, stability, and parsimony.
• The study provides a framework for developing reliable and interpretable models for biological discovery.