This paper presents a Genetic Algorithm, called Olex-GA, for the induction of rule-based text classifiers of the form ``classify document $d$ under category $c$ if $t_1 in d$ or ... or $t_n in d$ and not ($t_{n+1} in d$ or ... or $t_{n+m} in d$) holds'', where each $t_i$ is a term. Olex-GA relies on an efficient emph{several-rules-per-individual} binary representation and uses the $F$-measure as the fitness function.

The proposed approach is tested over the standard test sets Reuters and Ohsumed and compared against several classification algorithms (namely, Naive Bayes, Ripper, C4.5, SVM). Experimental results demonstrate that it achieves very good performance on both data collections, showing to be competitive with (and indeed outperforming in some cases) the evaluated classifiers.

Note:

A prototype of the rule induction system Olex-GA described in that paper is available at the address http://www.mat.unical.it/Olex-GA