Dynamic Selection of Classifiers Applied to High-Dimensional Small-Instance Data Sets: Problems and Challenges

Abstract

Dynamic selection (DS) of classifiers have been explored by researchers due to their overall ability to obtain higher accuracy on low-sample data sets when compared majority voting. Little literature, however, has employed DS to high-dimensional data sets with substantially more features than samples. Since, several studies have reported the benefits of applying feature selection methods to high-dimensional data sets, raised the following open research questions: 1. How DS methods perform for such data sets? 2. Do they perform better than majority voting? and 3. Does feature selection as a pre-processing step improve their performance? The performance of 21 DS methods was statistically compared against the performance of majority voting on 10 high-dimensional data sets and with a filter feature selection method. We found that majority voting is among the best ranked classifiers and none of the DS methods perform statistically better than it with and without feature selection. Moreover, we demonstrated that feature selection does improve the performance of DS methods.