Massive Pruning for Building an Operational Set of Association Rules: Metarules for Eliminating Conflicting and Redundant Rules

Extracting a set of association rules (AR) is a common method for representing knowledge embedded in a database. As long as many authors have aimed at improving the individual quality of these rules, not so many have considered their global quality and cohesiveness: Our objective is to provide the user with a set of rules he/she may combine to reason with, a consistent set as regards to "common sense logic". As local quality measures offer no warranty in this respect, we have defined patterns of major incoherencies and have associated metarules to them, resulting in a post-treatment cleaning phase for tracking down incoherencies and proposing corrections. We show that on the artificial Lucas0 database of the Causality Challenge, starting from 100 000 rules, we have reduced this rule set by three orders of magnitude, to 69 high-quality condensed rules embedding most of the structure designed by the challenge organizers.