Information Transmission and Nonspecificity in Feature Selection

Abstract

In this paper we propose a novel feature selection method which is based on fuzzy measures. More specifically, we apply a similarity measure to form similarity matrices from the data and apply nonspecificity on similarity degrees in order to conduct feature selection. To measure how relevant a particular feature is, we apply an information transmission measure. We exemplify our method on a simple artificial case to demonstrate its ability to select informative features. Moreover, we test our method on two real world data sets, the chronic kidney disease and the diabetic retinopathy Debrecen dataset. The nonspecificity-based feature selection method leads for both datasets to improvements in the mean classification performance. In comparison with the popular ReliefF algorithm and the Fisher Score, the new method reaches competitive results and also accomplishes the highest mean accuracy for both datasets.