DescriptionLogical Analysis of Data (LAD) is a machine learning/data mining methodology that combines ideas from areas such as Boolean functions, optimization and logic. In this thesis, we focus on the description and the application of novel optimization models to the construction of improved and/or simplified LAD models of data. We address the construction of LAD classification models, proposing two alternative ways of generating patterns, or rules. First, we show how to construct LAD models based on patterns of maximum coverage. We show, through a series of computational experiments, that such models are as good as, if not better than those obtained with the standard LAD implementation and other machine learning methods, while requiring a much simpler calibration for optimal performance. We formulate the problem of finding the most suitable LAD model as a large linear program, and show how to solve it using column generation. For the subproblem phase, we describe a branch-and-bound algorithm, whose performance is significantly superior to that of a commercial integer programming solver. The
LAD models produced by this algorithm are virtually parameter-free and practically as accurate as the calibrated models obtained with other machine learning methods. Finally, we propose a novel regression algorithm that extends the LAD methodology for the case of a numerical outcome and show that it constitutes an attractive alternative to other regression methods in terms of performance and flexibility of use.