Feature selection via modified RSBRA for SVM classifiers

Discretization can remove redundant and irrelative attributes during converting continuous attributes into discretized ones and therefore can be used for feature selection. Rough sets and Boolean reasoning based discretization approach (RSBRA), put forward by Nguyen in 1995, is very noticeable for its efficiency of reduction. However, the RSBRA is not a suitable feature selection method for machine learning algorithm such as neural network or SVM because too much useful information are lost due to the discretization. In this paper, we present a modified RSBRA for feature selection and evaluate it with SVM classifiers. In the presented algorithm, the level of consistency, coined from the rough sets theory, is introduced to substitute the stop criterion of circulation of the RSBRA, which maintains the fidelity of the training set after discretization. Experiment results show the modified algorithm has better predictive accuracies and less training time than the original RSBRA.