Sure Screening with Kernel-Based Distance Correlation: Methodology and Applications

Abstract

We consider a generalized kernel-based distance correlation measure for feature screening in high-dimensional settings. Theoretical results establish the sure screening property under mild regularity conditions for a class of negative-definite kernels. The method is flexible, requiring minimal distributional assumptions, and can be naturally extended to multivariate responses and grouped features. Extensive simulation studies confirm its robustness and effectiveness, while applications to real-world biomedical datasets demonstrate its practical relevance. The results highlight the potential of kernel-based distance measures as a powerful and scalable tool for variable selection in complex data environments.