A two-phase feature selection method using both filter and wrapper

This work proposes a two-phase (filter and wrapper) feature selection algorithm that can take advantage of both approaches to find an optimal subset of features.

Abstract

Feature selection is an integral step of the data mining process to find an optimal subset of features. After examining the problems with both the filter and the wrapper approach to feature selection, we propose a two-phase (filter and wrapper) feature selection algorithm that can take advantage of both approaches. It begins by running GFSIC (Genetic Feature Selection with Inconsistency Criterion), a filter approach, to remove irrelevant features, then it runs SBFCV (Sensitivity-Based Feature selection with v-fold Cross-Validation), a wrapper approach, to remove redundant or useless features. Analysis and experimental studies show the effectiveness and scalability of the proposed algorithm. The generalization of the neural network is improved when the algorithm is used to pre-process the training data by eliminating irrelevant and useless features from the neural network's consideration.