Micheline Kamber, Rajjan Shinghal
Knowledge discovery systems can be used to generate rules describing data from databases. Typically, only a small fraction of the rules generated are of interest. Measures of rule interestingness are hence essential for filtering out useless information. Such measures have been predominantly objective, based on statistics underlying the discovered rules, or patterns. Examples include the J-measure, rule strength, and certainty. Although these measures help assess the interestingness of discriminant rules, they do not fully serve their purpose when applied to characteristic rules. Discriminant rules describe how objects of a class differ from objects of other classes. We propose an interestingness measure for characteristic rules, based on the technical definition of sufficiency.