On kernel-based dependence measures for variable selection

Abstract

We investigate a generalized kernel-based distance correlation framework for feature screening in high-dimensional data. Under mild and interpretable conditions, and for a broad class of negative-definite kernels, we establish theoretical guarantees for the sure screening property. The flexibility of the proposed approach is demonstrated through an extensive empirical study covering multiple data types. Simulation results illustrate its robustness and efficiency, while applications to real biomedical datasets confirm its practical relevance. These results highlight kernel-based distance measures as a powerful tool for variable selection in complex data.