Applied Soft Computing Journal 93 (2020) 106402 Contents lists available at ScienceDirect Applied Soft Computing Journal journal homepage: www.elsevier.com/locate/asoc Feature selection based on improved binary global harmony search for data classification Jafar Gholami a , Farhad Pourpanah b , Xizhao Wang c ,* a Department of Computer Engineering, Kermanshah Science and Research Branch, Islamic Azad University, Kermanshah, Iran b College of Mathematics and Statistics, Shenzhen University, China c College of Computer Science and Software Engineering, Guangdong Key Lab. of Intelligent Information Processing, Shenzhen University, China article info Article history: Received 16 February 2020 Received in revised form 20 April 2020 Accepted 12 May 2020 Available online 22 May 2020 Keywords: Feature selection Population-based optimization Binary harmony search Data classification abstract Harmony search (HS) is an effective meta-heuristic algorithm inspired by the music improvisation process, where musicians search for a pleasing harmony by adjusting their instruments’ pitches. The HS algorithm and its variants have been widely used to solve binary and continuous optimization problems. In this paper, we propose an improved binary global harmony search algorithm, called IBGHS, to undertake feature selection problems. A modified improvisation step is introduced to enhance the global search ability and increase the convergence speed of the algorithm. In addition, the K -nearest neighbor (KNN) is used as an underlying learning model to evaluate the effectiveness of the selected feature subsets. The experimental results on eighteen benchmark problems indicate that the proposed IBGHS algorithm is able to produce comparable results as compared with other state-of- the-art population-based methods such as genetic algorithm (GA), particle swarm optimization (PSO), antlion optimizer (ALO), novel global harmony search (NGHS) and whale optimization algorithm (WOA) in solving feature selection problems. © 2020 Elsevier B.V. All rights reserved. 1. Introduction Feature selection (FS) is an optimization problem that plays an important role in tackling classification problems. It is a process of selecting an optimal subset of features from a data set so that the classifier can obtain better accuracy and/or reduce the computational burden. Nonetheless, removing irrelevant features is a challenging issue and time consuming owing to a large search space and wrapped relationship between the features [1–4]. FS techniques can be grouped into filter-, embedded- and wrapper-based methods [5]. The filter-based methods use the properties of the learning samples, such as distance and similar- ity, during the FS process [6]. Embedded-based methods search for the best feature subset during the training process, in order to reduce the computational burden [7]. While, wrapper-based methods use a classification algorithm to evaluate the quality of the various feature subsets, and a search mechanism to find the optimal ones. Among them, wrapper-based methods are more effective since they use a classifier to operate as a feedback mechanism to compute the fitness value of the selected feature subsets, but they are computationally expensive [8]. * Corresponding author. E-mail addresses: gholami2018@gmail.com (J. Gholami), farhad@szu.edu.cn (F. Pourpanah), xizhaowang@ieee.org (X. Wang). Traditional wrapper-based FS methods, such as sequential backward selection (SBS) [9] and sequential forward selection (SFS) [10], improve the performance of the learning model via sequentially adding or removing features from data set. In these methods, once features are removed or added, they cannot be updated in the next steps. Later, this problem was solved by inte- grating a floating technique into SBS and SFS [11]. However, they suffer from nesting effects and computationally expensive [12]. To alleviate these problems, population-based optimization al- gorithms, such as particle swarm optimization (PSO) [13–15], genetic algorithm (GA) [16–18], genetic programming (GP) [19, 20], ant colony optimization (ACO) [21], brain storm optimization (BSO) [22,23] and harmony search (HS) [24], have been used. These algorithms start with a set of randomly generated solu- tions, and use a fitness function to evaluate them. Then, they generate new solutions based on the individuals that performed better in the previous iteration. As a result, these algorithms reduce the computational burden as they avoid generating new individuals similar to the low quality ones. Among them, harmony search (HS) [25] is an effective meta- heuristic algorithm inspired by the music improvisation process of probing for a better state of harmony. HS has been widely ap- plied to solve real-world optimization problems, such as control system [26] and financial management [27], due to its simple structure, easy to implement and less parameters [28]. However, the basic HS algorithm suffers from several limitations such as https://doi.org/10.1016/j.asoc.2020.106402 1568-4946/© 2020 Elsevier B.V. All rights reserved.