The EFS-Server: A Web-Application for Feature Selection in Binary Classification

Feature selection methods are essential to identify a subset of features that improve the prediction performance of subsequent classification models and thereby also simplify their interpretability. Preceding studies showed the defectiveness in terms of specific biases of single feature selection methods, whereas an ensemble of feature selection techniques has the advantage to alleviate and compensate for such biases. With the development of the ensemble feature selection (EFS) method we take advantage of the benefits of multiple feature selection methods and combine their normalized outputs to a quantitative ensemble importance. Eight different feature selection methods have been used for the EFS approach. We evaluated the EFS method on a testset and it turned out that the subset of features retrieved by the EFS method showed a significantly improved performance in a subsequent logistic regression (LR) model compared to a model using all available features. EFS can be downloaded as an R-package or used in a websever at http://EFS.heiderlab.de.