A novel hybrid feature selection method considering feature interaction in neighborhood rough set

Abstract

Abstract The interaction between features can provide essential information that affects the performances of learning models. Nevertheless, most feature selection methods do not take interaction into account in feature correlations calculation. In this work, to solve the problem of dimensional reduction in hybrid data with uncertainty and noise, a novel feature selection method is proposed considering the characteristic of interaction in the neighborhood rough set. First of all, the multi-neighborhood radii set for hybrid data is obtained according to the distribution characteristics of features. Then, considering the ubiquity of interactive features, the feature correlations are redefined via employing various neighborhood information uncertainty measures. Furthermore, a new objective evaluation function of the interactive selection of hybrid features is developed, which is called the Max-Relevance min-Redundancy Max-Interaction (MRmRMI). Finally, a novel interaction feature selection algorithm based on neighborhood conditional mutual information (NCMI_IFS) is designed. To evaluate the performance of the proposed algorithm, we compare it with other eight representative feature selection algorithms on twenty public datasets. Experimental results on four different classifiers show that the NCMI_IFS algorithm has higher classification performance and is significantly effective.