Rule Learning with Negation for Text Classification

Classification rule generators that have the potential to include negated features in their antecedents are generally acknowledged to generate rules that have greater discriminating power than rules without negation. This can be achieved by including the negation of all features as part of the input. However, this is only viable if the number of features is relatively small. There are many applications where this is not the case, for example text classification. Given a large number of features, alternative strategies for including negated features in rules are desirable. This paper explores a number of strategies whereby this can be achieved in the context of inductive rule learning. Eight different strategies are proposed and evaluated by comparison with JRip, NaiveBayes and SMO. The reported results demonstrate that rules with negation produced a better classifier and that our rule learning mechanism outperform JRip and NaiveBayes and is competitive with SMO.