Feature selection for support vector regression using a genetic algorithm.

Area of Science:

• Machine Learning
• Bioinformatics
• Statistical Modeling

Background:

• Support vector regression (SVR) excels with nonlinear relationships but risks overfitting with numerous covariates.
• High dimensionality can reduce SVR's predictive accuracy due to overfitting.
• Effective feature selection is crucial for optimizing SVR performance in complex datasets.

Purpose of the Study:

• To develop and evaluate a genetic algorithm-based feature selection method for SVR.
• To enhance predictive accuracy of SVR by identifying optimal covariate subsets.
• To compare the proposed method against LASSO and random forest for SVR feature selection.

Main Methods:

• A genetic algorithm iteratively searches for optimal covariate subsets to maximize SVR performance.
• Performance was evaluated using a simulation study comparing the proposed method with SVR alone, LASSO, and random forest.
• The method was applied to predict 1-year donor kidney function using United Network for Organ Sharing registry data.

Main Results:

• The genetic algorithm feature selection method significantly improved SVR predictive accuracy compared to SVR without feature selection.
• The proposed method outperformed LASSO, particularly when nonlinear relationships between covariates and the outcome were present.
• Random forest showed comparable performance in some cases but was less effective with correlated covariates.

Conclusions:

• Genetic algorithm-based feature selection is a robust strategy for enhancing SVR predictive performance, especially in high-dimensional and nonlinear settings.
• This method offers a valuable alternative to existing techniques like LASSO and random forest for SVR optimization.
• The approach demonstrates practical utility in predicting clinical outcomes, such as donor kidney function post-transplant.