International Journal of Computer Theory and Engineering

Abstract

Feature selection (FS) is a key step in the data mining process. In FS, the objective is to select the smallest subset of features that reduces complexity and ensures generalization. In this paper, we present a combined filter-wrapper feature selection approach using misclassified data. The learning process starts with only one feature, which gives a large number of misclassified patterns. Only these patterns are used to select the next best feature which is added to the first one. By focusing on the misclassified patterns, the learner is undistracted and hence, it can select the relevant features more effectively and faster. The process continues until the classification results are within the required accuracy. The approach is applied to three datasets with high dimensional features using a variety of selection models and search strategies. Experimental results demonstrate the efficiency of the proposed approach in the two-class classification tasks.