G ift : granularity over specific-class for feature selection

As a fundamental material of Granular Computing, information granulation sheds new light on the topic of feature selection. Although information granulation has been effectively applied to feature selection, existing feature selection methods lack the characterization of feature potential. Such an ability is one of the important factors in evaluating the importance of features, which determines whether candidate features have sufficient ability to distinguish different target variables. In view of this, a novel concept of granularity over specific-class from the perspective of information granulation is proposed. Essentially, such a granularity is a fusion of intra-class and extra-class based granularities, which enables to exploit the discrimination ability of features. Accordingly, an intuitive yet effective framework named G ift , i.e., granularity over specific-class for feature selection, is proposed. Comprehensive experiments on 29 public datasets clearly validate the effectiveness of G ift as compared with other feature selection strategies, especially in noisy data.