SOM+PSO

Abstract

Currently, most processes have a volume of historical information that makes its manual processing difficult. Data mining, one of the most significant stages in the Knowledge Discovery in Databases (KDD) process, has a set of techniques capable of modeling and summarizing these historical data, making it easier to understand them and helping the decision making process in future situations. This article presents a new data mining adaptive technique called SOM+PSO that can build, from the available information, a reduced set of simple classification rules from which the most significant relations between the features recorded can be derived. These rules operate both on numeric and nominal attributes, and they are built by combining a variation of a population metaheuristic and a competitive neural network. The method proposed was compared with the PART method and measured over 19 databases (mostly from the UCI repository), and satisfactory results were obtained.