Neurocomputing 306 (2018) 94–107 Contents lists available at ScienceDirect Neurocomputing journal homepage: www.elsevier.com/locate/neucom Novel multiobjective TLBO algorithms for the feature subset selection problem Hakan Ezgi Kiziloz a,∗ , Ayça Deniz b , Tansel Dokeroglu a , Ahmet Cosar b a Computer Engineering Department, Turkish Aeronautical Association University, Ankara, Turkey b Computer Engineering Department, Middle East Technical University, Ankara, Turkey a r t i c l e i n f o Article history: Received 9 September 2017 Revised 11 March 2018 Accepted 20 April 2018 Available online 4 May 2018 Communicated by W K Wong Keywords: Teaching learning based optimization Multiobjective feature selection Supervised learning a b s t r a c t Teaching Learning Based Optimization (TLBO) is a new metaheuristic that has been successfully applied to several intractable optimization problems in recent years. In this study, we propose a set of novel mul- tiobjective TLBO algorithms combined with supervised machine learning techniques for the solution of Feature Subset Selection (FSS) in Binary Classification Problems (FSS-BCP). Selecting the minimum num- ber of features while not compromising the accuracy of the results in FSS-BCP is a multiobjective opti- mization problem. We propose TLBO as a FSS mechanism and utilize its algorithm-specific parameterless concept that does not require any parameters to be tuned during the optimization. Most of the classical metaheuristics such as Genetic and Particle Swarm Optimization algorithms need additional efforts for tuning their parameters (crossover ratio, mutation ratio, velocity of particle, inertia weight, etc.), which may have an adverse influence on their performance. Comprehensive experiments are carried out on the well-known machine learning datasets of UCI Machine Learning Repository and significant improvements have been observed when the proposed multiobjective TLBO algorithms are compared with state-of-the- art NSGA-II, Particle Swarm Optimization, Tabu Search, Greedy Search, and Scatter Search algorithms. © 2018 Elsevier B.V. All rights reserved. 1. Introduction With the recent improvements in science and technology, huge amounts of data is being generated everyday. The size of data is larger than a human can process without help of an intelligent sys- tem [1]. This exploding growth of data makes researchers search for new methods to extract meaningful information. Effective decision-making requires high quality in information/knowledge [2]. However, it becomes harder to extract meaningful information as the amount of raw input data increases. If the raw input data is not preprocessed (e.g. filtering), it may have adverse effects and mislead the decision making processes. This creates a rapidly increasing demand for advanced data processing techniques such as data mining and machine learning. Data mining identifies the existing patterns that might help predict future behaviours. In addition to data mining techniques, machine learning techniques are also widely used in modern decision making process. Data mining modifies data by filtering, formatting, etc., whereas machine learning techniques benefit ∗ Corresponding author. E-mail addresses: hakanezgi@etu.edu.tr (H.E. Kiziloz), ayca.deniz@metu.edu.tr (A. Deniz), tansel.dokeroglu@thk.edu.tr (T. Dokeroglu), cosar@ceng.metu.edu.tr (A. Cosar). from historical data to build a smart model [3]. Large amounts of data can be analyzed in a limited time by using machine learning techniques. Researchers agree on the fact that preprocessing enables data mining tools to perform more effectively [4]. One of the most commonly applied data preprocessing techniques is Feature Subset Selection (FSS), which is the process of reducing the number of features by identifying irrelevant or redundant attributes of a dataset that do not affect or make no contribution to the solution of the problem [5]. However, in the meantime, we should mini- mize any loss of critical information. Machine learning algorithms will, naturally, execute faster when the amount they process is de- creased by using FSS. The accuracy of the results may also improve in some cases [6]. As data grow massively, FSS becomes indispens- able in order to be able to extract meaningful information. FSS algorithms are widely applied in various real-world problems such as text categorization and recommendation systems [7–9]. FSS is a multiobjective optimization process with two objec- tives, maximizing the accuracy of the results and minimizing the number of features. Therefore, there can be a set of solutions rather than a single one. The set of solutions serves both objec- tives and cannot dominate each other. For example, a solution may have an accuracy value of 0.85 with five features whereas another solution may have an accuracy value of 0.75 with three https://doi.org/10.1016/j.neucom.2018.04.020 0925-2312/© 2018 Elsevier B.V. All rights reserved.