Generating a set of association and decision rules with statistically representative support and anti-support

The problem of rule evaluation is central to association and decision rule mining. In the literature, many attractiveness measures assessing the utility of discovered rules have been proposed. In particular, two-dimensional evaluation planes like support - confidence or support - anti-support are a frequently employed scheme. The inherent drawback of attractiveness measures is the necessity to provide a cutoff threshold which defines the minimal or maximal acceptable value of a given measure. Such threshold is usually unintuitive, meaningless and difficult to impose with objectivity. In this paper, we focus on the support - anti-support evaluation plane. We propose a methodology of simultaneously assessing statistical representativeness of both measures by performing multinomial tests on association or decision rules. The proposed approach combines assessing statistical soundness of the rules with an implicit eliciting of a threshold in both measures from the user. The latter is accomplished by employing the notion of a relative error in both support and anti-support. We evaluate the proposed method on a number of data sets and provide general conclusions.