A Simulation Technique for extracting Robust Association Rules
Résumé
Data mining techniques like association rules extraction can be used to analyse large amounts of data (Han et al. 2001). To make sure they produce interesting knowledge, researchers have already defined more than fifty quality measures for association rules. These measures are constructed to sort rules with different semantic approaches (Briand et al. 2004). In this paper, our aim is not to order rules extracted from data but to differenciate accidental rules from generalizable rules. If data are in a table of type objectsXattributes, with numerous objects, few attributes, filled with the quantitative values of these objects to these attributes, creation of samples in statistics allows to separate accidental results from generalizable ones. But statistics work badly with numerous boolean attributes (Cadot et al. 2005). In this paper, we present a simulation technique to determine if the links between these numerous boolean attributes expressed by frequent itemsets are accidental or not, with a risk error less than a fixed value (for example, alpha=.05). We generate numerous (for example a hundred) artificial data with the same statistical properties as real data, except the links between attributes. So we compute frequency thresholds which separate accidental itemsets to significant itemsets. This technique is compared with acceptation/reject techniques and applied to real data.