Feature selection based on maximal neighborhood discernibility

Neighborhood rough set has been proven to be an effective tool for feature selection. In this model, the positive region of decision is used to evaluate the classification ability of a subset of candidate features. It is computed by just considering consistent samples. However, the classification ability is not only related to consistent samples, but also to the ability to discriminate samples with different decisions. Hence, the dependency function, constructed by the positive region, cannot reflect the actual classification ability of a feature subset. In this paper, we propose a new feature evaluation function for feature selection by using discernibility matrix. We first introduce the concept of neighborhood discernibility matrix to characterize the classification ability of a feature subset. We then present the relationship between distance matrix and discernibility matrix, and construct a feature evaluation function based on discernibility matrix. It is used to measure the significance of a candidate feature. The proposed model not only maintains the maximal dependency function, but also can select features with the greatest discernibility ability. The experimental results show that the proposed method can be used to deal with heterogeneous data sets. It is able to find effective feature subsets in comparison with some existing algorithms.