European Journal of Operational Research 248 (2016) 593–606 Contents lists available at ScienceDirect European Journal of Operational Research journal homepage: www.elsevier.com/locate/ejor Computational Intelligence and Information Management A pool-based pattern generation algorithm for logical analysis of data with automatic fine-tuning Marco Caserta a,∗ , Torsten Reiners b a IE University and IE Business School, Maria de Molina, 31-B, 28006, Madrid, Spain b Curtin University, Bentley, WA, 6102, Australia article info Article history: Received 2 October 2013 Accepted 30 May 2015 Available online 6 June 2015 Keywords: Logical Analysis of Data Data mining Fine-tuning bRKGA Machine learning abstract In this paper, we address the binary classification problem, in which one is given a set of observations, char- acterized by a number of (binary and non-binary) attributes and wants to determine which class each ob- servation belongs to. The proposed classification algorithm is based on the Logical Analysis of Data (LAD) technique and belongs to the class of supervised learning algorithms. We introduce a novel metaheuristic- based approach for pattern generation within LAD. The key idea relies on the generation of a pool of patterns for each given observation of the training set. Such a pool is built with one or more criteria in mind (e.g., diversity, homogeneity, coverage, etc.), and is paramount in the achievement of high classification accuracy, as shown by the computational results we obtained. In addition, we address one of the major concerns of many data mining algorithms, i.e., the fine-tuning and calibration of parameters. We employ here a novel technique, called biased Random-Key Genetic Algorithm that allows the calibration of all the parameters of the algorithm in an automatic fashion, hence reducing the fine-tuning effort required and enhancing the performance of the algorithm itself. We tested the proposed approach on 10 benchmark instances from the UCI repository and we proved that the algorithm is competitive, both in terms of classification accuracy and running time. © 2015 Elsevier B.V. and Association of European Operational Research Societies (EURO) within the International Federation of Operational Research Societies (IFORS). All rights reserved. 1. Introduction Let us consider a binary classification problem, in which one is given a dataset composed of observations belonging to one of two classes, e.g., positive or negative, where the class each observation belongs to is known. A typical data mining problem is the classifica- tion problem, i.e., finding the class a new observation, not included in the dataset, belongs to. Binary classification finds a large number of applications, spanning from, e.g., medical diagnosis (Alexe, Alexe, Axelrod, Hammer, & Weissmann, 2005; Hammer & Bonates, 2006), to credit risk rating (Hammer, Kogan, & Lejeune, 2006), from main- tenance replacement (Ghasemi & Esameili, 2013), to fault diagnosis (Mortada, Yacout, & Lakis, 2013). Owing to the fact that classification is such a relevant problem in the data mining field, effective techniques to classify data have been developed. The observations to be classified are characterized by a set of attributes, which are believed to affect the class each observa- tion belongs to. However, the relationship between the value of the ∗ Corresponding author. Tel.: +34 915689600. E-mail addresses: marco.caserta@ie.edu (M. Caserta), t.reiners@curtin.edu.au (T. Reiners). attributes and the class is unknown and, therefore, needs to be esti- mated via a training process. The classifier discovers such rules that map an object to a class, based on the values of the attributes. Logical Analysis of Data (LAD) was introduced in Boros, Hammer, Ibaraki, and Kogan (1997), and Boros, Hammer, Ibaraki, Kogan, Mayoraz, and Muchnik (2000) and is a data analysis methodology that combines ideas from combinatorial optimization and Boolean functions and belongs to the family of supervised learning tech- niques. The LAD methodology relies on a “rule learning” mechanism and, therefore, is strongly connected with other popular classifica- tion rules presented in the machine learning literature. According to Fürnkranz (1999), many rule learning algorithms are based on a se- quential covering procedure, in which the steps to follow are: “Learn a rule that covers part of the given training examples, remove the covered examples from the training set, and recursively learn another rule that covers some of the remaining examples until no examples remain.” LAD fits well within this sequential covering framework. For a more extensive discussion, we refer the reader interested in the connection between LAD and other “rule learning” techniques as well as on the “justifiability” of LAD to Boros, Crama, Hammer, Ibaraki, Kogan, and Makino (2011). Along the same line, Dembczynski, Kotlowski, and Slowinski (2010) establish a clear connection be- tween LAD and other methods that fall within the framework of http://dx.doi.org/10.1016/j.ejor.2015.05.078 0377-2217/© 2015 Elsevier B.V. and Association of European Operational Research Societies (EURO) within the International Federation of Operational Research Societies (IFORS). All rights reserved.