Portal de Periódicos da CAPES

IOGA: An instance-oriented genetic algorithm

1996; Springer Science+Business Media; Linguagem: Inglês

10.1007/3-540-61723-x_1012

1611-3349

Richard S. Forsyth,

Data Mining Algorithms and Applications

Instance-based methods of classification are easy to implement, easy to explain and relatively robust. Furthermore, they have often been found in empirical studies to be competitive in accuracy with more sophisticated classification techniques (Aha et al., 1991; Weiss & Kulikowski, 1991; Fogarty, 1992; Michie et al., 1994). However, a twofold drawback of the simplest instance-based classification method (1-NNC) is that it requires the storage of all training instances and the use of all attributes or features on which those instances are measured — thus failing to exhibit the cognitive economy which is the hallmark of successful learning (Wolff, 1991). Previous researchers have proposed ways of adapting the basic 1-NNC algorithm either to select only a subset of training cases (‘prototypes’) or to discard redundant and/or ‘noisy’ attributes, but not to do both at once. The present paper describes a program (IOGA) that uses an evolutionary algorithm to select prototypical cases and relevant attributes simultaneously, and evaluates it empirically by application to a set of test problems from a variety of fields. These trials show that very considerable economization of storage can be achieved, coupled with a modest gain in accuracy.