Abstract
Faced with the increase in high-dimensional Big Data creating more volume and complexity, the feature selection process became an essential phase in the preprocessing workflow upstream of the design of systems based on deep learning. This paper is a concrete and first application of the new metaheuristic Harris Hawk Optimization Encirclement-Attack-Synergy (HHO-EAS) in solving the NP-Hard wrapper feature selection multi-objective optimization problem. This problem combines two contradictory objectives: maximizing the accuracy of a classifier while minimizing the number of the most relevant and non-redundant selected features. To do this we hybridized HHO-EAS to create the new metaheuristic Binary HHO-EAS (BHHO-EAS). We combined HHO-EAS to the sixteen transfer functions most used in the literature structured in a balanced way among the four main categories including S-Shaped, V-Shaped, Q-Shaped and U-Shaped. This wide range of transfer function allows us to analyze the evolution of BHHO-EAS’s skills according to the assigned transfer function and to determine which of them offer the best performances. We applied wrapper feature selection to the well-known NSL-KDD dataset with the deep learning Multi Layer Perceptron (MLP) classifier. We put BHHO-EAS in competition with three other well-known population based binary metaheuristics, BPSO, BBA and BHHO. The analysis of the experimental results, compared to the three other binary metaheuristics, demonstrated that BHHO-EAS obtained the best performance on 100% of the transfer functions. This is more particularly highlighted by the U-Shaped transfer functions, which give an acceptable compromise for the two objectives of the problem with an average accuracy of 96,4% and an average size of selected features of 20.