一种基于ReliefF评估和互补系数的特征选择算法

Filter特征选择算法具有通用性强、算法复杂度低的特点,但对某一个具体的分类器选择的特征子集也许并不是最优的;Wrapper方法与其相反,对特定的分类器可以找到最优的特征子集,但算法复杂度很高.研究一种Filter与Wrapper相结合的混合型算法.首先从特征对样本分类效果的角度提出互补系数的概念,然后基于ReliefF评估和互补系数,提出ReCom算法.实验证明,由ReCom算法得到的特征子集与ReliefF算法得到的特征子集相比具有更好的性能,并且与传统Wrapper方法相比,该算法大大降低了时间复杂度.

A feature selection method based on ReliefF evaluation and complementary coefficient

Filter feature selection algorithm has the advantages of low complexity and being adaptive to different databases, the drawback is that the selected feature subset may not be the best for a given classifier. On the contrary, Wrapper method can select the best feature subset for a given classifier, but has high complexity. In this paper, a method combining Filter with Wrapper is stuided. First, complementary coefficient is presented to measure the correlation between features. Then, ReCom algorithm is presented based on ReliefF evaluation and complementary coefficient. Experimental results show that ReCom algorithm has better performance than ReliefF algorithm, and it has lower time complexity than general Wrapper algorithms.