Abstract
Feature selection is an important pre-processing step in classification problems. It can reduce the dimensionality of a dataset and increase the accuracy and efficiency of a learning/classification algorithm. Filter methods are necessary to obtain only the relevant features to the class and to avoid redundancy. While wrapper methods are applied to get optimized features and better classification accuracy. This paper proposes a feature selection based on hybridization of mutual information feature selection (MIFS) filter and modified binary cuckoo search (MBCS) wrapper methods. The classifier accuracy of K-nearest neighbor (KNN) is used as the fitness function. The experimental results show that the hybrid filter-wrapper algorithm maintains the high classification performance achieved by wrapper methods and significantly reduce the computational time. At the same time, it reduces the number of features.