A Novel Hybrid Method for Feature Subset Selection

N. Snchez-Maroo, A. Alonso-Betanzos, and M. Tombilla-Sanrom n (Spain)


Machine learning, feature selection, CFS filter, IAFN-FS wrapper


In this paper, a new hybrid method for feature selection is proposed. Feature selection has become a commonly tech nique used in machine learning because it may improve ac curacy and efficiency of classifier methods. The method proposed in this paper combines the two main approaches for feature selection: filters and wrappers. This is due to the fact that, although wrapper models tend to give better results than filters, the latter are computationally less ex pensive. This is important for problems with a high num ber of input characteristics. Specifically, the first step of the method consists on applying a Correlation-based Feature Selection (CFS) filter. Subsequently, a wrapper method, In cremental ANOVA Functional Networks-Feature Selection (IAFN-FS), is applied over the features selected by the fi ter. The performance of the hybrid method is tested using several benchmark data sets and the results are compared to those obtained by wrapper methods that combine well known induction algorithms such as Naive Bayes and C4.5 with backward and forward search strategies.

Important Links:

Go Back