A comparison of different data transformation approaches in the feature ranking context

Abstract

Due to the omnipresence of high-dimensional datasets, feature selection and ranking are very important steps in data preprocessing. In this work, we propose three transformations for real-valued features. The transformations are based on estimating the probability densities of the features. Originally, we propose modified distance measures for the ReliefF algorithm, which is one the most prominent feature ranking algorithms. To enable their comparison with the other feature ranking algorithms, we present data transformations that are mathematically equivalent to the modified distance measures. Finally, we evaluate our proposed transformations used in combination with several feature ranking methods on a set of benchmark datasets.