Intl. Trans. in Op. Res. 25 (2018) 1027–1052 DOI: 10.1111/itor.12428 INTERNATIONAL TRANSACTIONS IN OPERATIONAL RESEARCH Interactive evolutionary approaches to multiobjective feature selection M¨ uberra ¨ Ozmen a ,G¨ uls ¸ah Karakaya b and Murat K ¨ oksalan a a Department of Industrial Engineering, Middle East Technical University, 06800 Ankara, Turkey b Department of Business Administration, Middle East Technical University, 06800 Ankara, Turkey E-mail: muberra.ozmen@gmail.com [ ¨ Ozmen]; kgulsah@metu.edu.tr [Karakaya]; koksalan@metu.edu.tr [K¨ oksalan] Received 30 September 2016; received in revised form 20 April 2017; accepted 24 April 2017 Abstract In feature selection problems, the aim is to select a subset of features to characterize an output of interest. In characterizing an output, we may want to consider multiple objectives such as maximizing classification performance, minimizing number of selected features or cost, etc. We develop a preference-based approach for multiobjective feature selection problems. Finding all Pareto-optimal subsets may turn out to be a compu- tationally demanding problem and we still would need to select a solution. Therefore, we develop interactive evolutionary approaches that aim to converge to a subset that is highly preferred by the decision maker (DM). We test our approaches on several instances simulating DM preferences by underlying preference functions and demonstrate that they work well. Keywords: feature selection; subset selection; interactive approach; evolutionary algorithm 1. Introduction In classification problems, supervised learning algorithms, such as decision trees, support vector machines (SVMs), neural networks, etc. are used to predict the class (or output variable) of an instance by observing its feature (or input variables) values. Supervised learning algorithms train a prediction model over a dataset, in which different feature and class values of some past observa- tions are provided, by understanding the relationship between the features and classes. Hence, the prediction model can be used to classify a new instance based on its features. The classification performance of a learning algorithm depends on its ability to detect the rela- tionship between input and output variables accurately. However, the presence of features that are irrelevant to the class, or the redundancy within the features, may have a negative impact on the classification performance of the learning algorithm (Kohavi and John, 1997). Yuand Liu (2004) classify the features based on their relevance with respect to the output as strongly relevant, weakly C  2017 The Authors. International Transactions in Operational Research C  2017 International Federation of Operational Research Societies Published by John Wiley & Sons Ltd, 9600 Garsington Road, Oxford OX4 2DQ, UK and 350 Main St, Malden, MA02148, USA.