Class-specific mutual information variation for feature selection

Feature selection plays a critical role in pattern recognition. Feature selection aims to eliminate irrelevant and redundant features. A drawback of traditional feature selection methods is that they ignore the dynamic change of selected features with the class. To address this problem, we develop a novel linear feature selection method, namely, Dynamic Change of Selected Feature with the class (DCSF). In DCSF, we introduce a new term: the conditional mutual information between the selected features and the class when a candidate feature is considered. In addition, we replace the traditional feature relevancy term with a term that is based on conditional mutual information. To evaluate our method, we compare DCSF with five traditional methods and two state-of-the-art methods on 20 benchmark data sets. Experimental results show that DCSF outperforms seven other methods in terms of average classification accuracy and highest classification accuracy.