Multiobjective particle swarm optimization in classification-rule learning

Abstract

The chapter presents a convenient algorithm of classification-rule learning that uses the Multiple Objective Particle Swarm Optimization technique to obtain a set of the best classification rules as an alternative way to cover algorithms, avoiding the loss of quality of the rules. The best rules are memorized in an unordered classifier, based on the concept of Pareto dominance, which permits the selection of rules with several characteristics that can be defined by the user. The rules are selected from the classifier at the same time that they are discovered and, differently from the cover algorithms, the selection does not remove examples from the dataset. We study classification-rule learning, describing some paradigms, different rule's representations and different particle initialization procedures. Then, we describe the way the particles move in the search space and the fitness functions. After that, we show some aspects of the required global repositories and describe aspects of different stopping criteria. Then, we present the proposed algorithm MOPSO-RL, showing its complexity and restrictions. Comparisons with other algorithms from the literature show that the proposed algorithm is competitive related to the area under ROC (Receiver Output Curve) and regarding the number of rules in the classifier. Besides, the selected rules have high support and weighted relative accuracy (Wracc), which denotes that the rules are important even when considered in isolation. In this way, it produces a set that is good for classification, with very good rules that can bring knowledge even if they are analyzed in isolation. © 2009 Springer-Verlag Berlin Heidelberg.