High-Dimensional Feature Selection by Feature-Wise Kernelized Lasso

摘要: The goal of supervised feature selection is to find a subset input features that are responsible for predicting output values. least absolute shrinkage and operator (Lasso) allows computationally efficient based on linear dependency between In this letter, we consider feature-wise kernelized Lasso capturing nonlinear input-output dependency. We first show with particular choices kernel functions, nonredundant strong statistical dependence values can be found in terms kernel-based independence measures such as the Hilbert-Schmidt criterion. then globally optimal solution efficiently computed; makes approach scalable high-dimensional problems. effectiveness proposed method demonstrated through experiments classification regression thousands features.