Fuzzy rule selection by multi-objective genetic local search algorithms and rule evaluation measures in data mining

✍ Scribed by Hisao Ishibuchi; Takashi Yamamoto

• Publisher: Elsevier Science
• Year: 2004
• Tongue: English
• Weight: 365 KB
• Volume: 141
• Category: Article
• ISSN: 0165-0114
• DOI: 10.1016/s0165-0114(03)00114-3

✦ Synopsis

This paper shows how a small number of simple fuzzy if-then rules can be selected for pattern classiÿcation problems with many continuous attributes. Our approach consists of two phases: candidate rule generation by rule evaluation measures in data mining and rule selection by multi-objective evolutionary algorithms. In our approach, ÿrst candidate fuzzy if-then rules are generated from numerical data and prescreened using two rule evaluation measures (i.e., conÿdence and support) in data mining. Then a small number of fuzzy if-then rules are selected from the prescreened candidate rules using multi-objective evolutionary algorithms. In rule selection, we use three objectives: maximization of the classiÿcation accuracy, minimization of the number of selected rules, and minimization of the total rule length. Thus the task of multi-objective evolutionary algorithms is to ÿnd a number of non-dominated rule sets with respect to these three objectives. The main contribution of this paper is to propose an idea of utilizing the two rule evaluation measures as prescreening criteria of candidate rules for fuzzy rule selection. An arbitrarily speciÿed number of candidate rules can be generated from numerical data for high-dimensional pattern classiÿcation problems. Through computer simulations, we demonstrate that such a prescreening procedure improves the e ciency of our approach to fuzzy rule selection. We also extend a multi-objective genetic algorithm (MOGA) in our former studies to a multi-objective genetic local search (MOGLS) algorithm where a local search procedure adjusts the selection (i.e., inclusion or exclusion) of each candidate rule. Furthermore, a learning algorithm of rule weights (i.e., certainty factors) is combined with our MOGLS algorithm. Such extensions to our MOGA for fuzzy rule selection are another contribution of this paper.