Model-free feature screening for ultrahigh dimensional classification

In this paper, a new model-free feature screening method based on classification accuracy of marginal classifiers is proposed for ultrahigh dimensional classification. Different from existing methods, which use the differences of means or differences of conditional cumulative distribution functions between classes as the screening indexes, we propose a new feature screening method to rank the importance of predictors based on classification accuracy of marginal classifiers. For each variable, we construct the corresponding marginal classifier according to the Bayes rule and thus classification accuracy of these marginal classifiers can be used as effective feature screening indexes to select all important variables. Not only for a fixed number of classes but also for a diverging number of classes, we can prove that the proposed method enjoys the sure screening property under some regularity conditions. Finally, simulations and the real data analysis well demonstrate good performance of the proposed method in comparison with existing methods.