A new probabilistic measure of interestingness for association rules, based on the likelihood of the link

Abstract

The interestingness measures for pattern associations proposed in the data mining literature depend only on the observation of relative frequencies obtained from 2 × 2 contingency tables. They can be called "absolute measures". The underlying scale of such a measure makes statistical decisions difficult. In this paper we present the foundations and the construction of a probabilistic interestingness measure that we call likelihood of the link index. This enables to capture surprising association rules. Indeed, its underlying principle can be related to that of information theory philosophy; but at a relational level. Two facets are developed for this index: symmetrical and asymmetrical. Two stages are needed to build this index. The first is "local" and associated with the two single boolean attributes to be compared. The second corresponds to a discriminant extension of the obtained probabilistic index for measuring an association rule in the context of a relevant set of association rules. Our construction is situated in the framework of the proposed indices in the data mining literature. Thus, new measures have been derived. Finally, we designed experiments to estimate the relevance of our statistical approach, this being theoretically validated, previously. © Springer-Verlag Berlin Heidelberg 2007.