A novel extended binary cuckoo search algorithm for feature selection

Abstract

Cuckoo Search is a recent nature-inspired metaheuristic algorithm, inspired by the cuckoo birds' aggressive strategy to breeding. The Cuckoo Search algorithm iteratively uses a Levy flight random walk to explore a search space. The Levy flight mechanism takes sudden turns of 90 degrees and consequently the Cuckoo's Search strategy does not carefully search around the cuckoos' nest, and hence it suffers from slow convergence and low optimisation accuracy. In order to improve the weaknesses of the Cuckoo Search algorithm, this paper proposes a pseudobinary mutation neighbourhood search procedure embedded in a new binary version of the Cuckoo Search algorithm. The proposed Extended Binary Cuckoo Search algorithm has been designed for the task of feature selection, and thus aims to minimise the number of selected features (such that only the best features are retained in the subset) whilst maximising the classification accuracy. Based on these criteria, a new objective function is proposed which considers the number of features in the subset as well as the classification accuracy when searching for the best subset of features. To measure the classification accuracy when using a set of candidate features, the Support Vector Machine classifier is utilised. Experiments were conducted with the proposed Extended Binary Cuckoo Search optimisation applied to biomedical datasets and the results demonstrated the superiority of the proposed algorithm against three other nature-inspired algorithms, namely the Binary Ant Colony Optimisation, Binary Genetic Algorithm, and Binary Particle Swarm Optimisation. Moreover, the experiments revealed that when using the proposed function, higher classification accuracy is achieved using a fewer number of features, as opposed to using the standard classification accuracy function which needed more features to achieve comparable accuracy.