Evaluation et validation des règles d'association
Stephane Lallich and Olivier Teytaud
Revue des nouvelles technologies de l'information, RNTI-E-1
Volume Vol 1,
The research of interesting asso ciations rules in databases is an imp ortant task of knowledge data discovery. Algorithms based on supp ort and condence, such as Apriori, brought a neat solution to the rules extraction problem. As shown in this article, these algorithms miss interesting rules and some of the rules they select are of no interest. Furthermore, they pro duce to o many rules. It is therefore imp ortant to have other measures to complement supp ort and condence. In this article, we review the dierent measures suggested in the literature and we prop ose criteria for their evaluation. We then suggest a validation
metho d using to ols issued from the statistical learning theory, notably VC-dimension.
Facing the high numb er of measures and the multitude of candidate rules, these to ols enable to set uniform non asymptotic b ounds for all rules and measures simultaneously.