Robust Variable Selection and Estimation Based on Kernel Modal Regression.

Area of Science:

• Machine Learning
• Statistical Modeling

Background:

• Model-free variable selection is increasingly popular for its flexibility.
• Existing methods often use the mean square error (MSE) criterion, which is sensitive to non-Gaussian noise and outliers.
• MSE-based methods struggle in complex environments due to assumptions of Gaussian noise.

Purpose of the Study:

• To develop a robust model-free variable selection algorithm.
• To overcome limitations of MSE-based methods in handling complex noise and outliers.
• To improve the effectiveness of model-free variable selection in diverse applications.

Main Methods:

• Integration of kernel modal regression and gradient-based variable identification.
• Utilizing a modal regression estimator linked to information theoretic learning and the maximum correntropy criterion.
• Replacing conditional mean learning with conditional mode learning for noise robustness.

Main Results:

• The proposed algorithm demonstrates robustness to complex noise and outliers.
• Gradient information from the estimator provides a model-free metric for variable screening.
• Theoretical analysis confirms generalization bounds and variable selection consistency.

Conclusions:

• The new algorithm offers a robust and effective approach to model-free variable selection.
• It addresses the limitations of traditional MSE-based methods in complex data scenarios.
• Empirical validation through data experiments confirms the method's practical effectiveness.