Efficient Top-K Feature Selection Using Coordinate Descent Method

Sparse learning based feature selection has been widely investigated in recent years. In this study, we focus on the l2,0-norm selection, which is effective for exact top-k but challenging to optimize. To solve general constrained problems, novelly develop a parameter-free optimization framework coordinate descend (CD) method, termed CD-LSR. Specifically, devise skillful conversion from original problem solving one continuous matrix and discrete matrix. Then nontrivial constraint can be solved efficiently by with CD method. We impose vanilla least square regression (LSR) model optimize it Extensive experiments exhibit efficiency of CD-LSR, as well discrimination ability identify informative features. More importantly, versatility CD-LSR facilitates applications more sophisticated models. Based competitive performance baseline LSR model, satisfactory its reasonably expected. The source MATLAB code are available at: https://github.com/solerxl/Code_For_AAAI_2023.