Using Data Complexity Measures for Thresholding in Feature Selection Rankers.

In the last few years, feature selection has become essential to confront the dimensionality problem, removing irrelevant and redundant information. For this purpose, ranker methods have become an approximation commonly used since they do not compromise the computational efficiency. Ranker methods return an ordered ranking of all the features, and thus it is necessary to establish a threshold to reduce the number of features to deal with. In this work, a practical subset of features is selected according to three different data complexity measures, releasing the user from the task of choosing a fixed threshold in advance. The proposed approach was tested on six different DNA microarray datasets which have brought a difficult challenge for researchers due to the high number of gene expression and the low number of patients. The adequacy of the proposed approach in terms of classification error was checked by the use of an ensemble of ranker methods with a Support Vector Machine as classifier. This study shows that our approach was able to achieve competitive results compared with those obtained by fixed threshold approach, which is the standard in most research works.