Adversarial Training Based Feature Selection

摘要

Feature selection is one of key problems in machine learning and data mining. It has been widely accepted that adversarial training is an effective strategy to improve the accuracy and robustness of classifiers. In this paper, in order to improve the performance of feature selection, adversarial training is also adopted, and an adversarial training based feature selection framework is proposed. To validate the effectiveness of the proposed feature selection framework, three classical feature selection algorithms, i.e. Relief-F, Fisher Score and minimum Redundancy and maximum Relevance (mRMR) are chosen and two methods are used to generate adversarial examples in experiments. The experimental results on benchmark datasets containing low-dimension and high-dimension datasets demonstrate show that adversarial training is able to improve the performance of classical feature selection methods in most cases.