An experimental comparison of feature selection methods on two-class biomedical datasets

Feature selection is a significant part of many machine learning applications dealing with small-sample and high-dimensional data. Choosing the most important features is an essential step for knowledge discovery in many areas of biomedical informatics. The increased popularity of feature selection methods and their frequent utilisation raise challenging new questions about the interpretability and stability of feature selection techniques. In this study, we compared the behaviour of ten state-of-the-art filter methods for feature selection in terms of their stability, similarity, and influence on prediction performance. All of the experiments were conducted on eight two-class datasets from biomedical areas. While entropy-based feature selection appears to be the most stable, the feature selection techniques yielding the highest prediction performance are minimum redundance maximum relevance method and feature selection based on Bhattacharyya distance. In general, univariate feature selection techniques perform similarly to or even better than more complex multivariate feature selection techniques with high-dimensional datasets. However, with more complex and smaller datasets multivariate methods slightly outperform univariate techniques.