Extraction de "pépites" de connaissance dans les données : une nouvelle approche et une étude de la sensibilité au bruit
Most of the classical approach for the extraction of association rules are based on the use of thresholds, set by the expert, to prune the search space. The choice of these thresholds, supposed to efficiently separate the set of interesting rules from the set of obvious rules, is quite difficult even for a domain expert. Considering that the data may be noisy and that the extraction of ``nuggets'' of knowledge (i.e., association rules with small support) may be of particular interest to the expert, then the classical methods are often unable to deal with this problem. We propose a new association rule extraction measure called ``least-contradiction''. We show that this measure (i) enables us to extract ``nuggets'' of knowledge from the data, whitout drowning in a huge amount of rules having small supports, (ii) reacts somewhat less badly to noise than the other classical measures.