Nonredundant Generalized Rules and Their Impact in Classification

François Rioult (1) , Bruno Zanuttini (2) , Bruno Crémilleux (1)

1 Equipe CODAG - Laboratoire GREYC - UMR6072
2 Equipe MAD - Laboratoire GREYC - UMR6072

Association rules are commonly used in classification based on associations. These rules are made of conjunctions of attributes in the premise and a class attribute in conclusion. In this chapter, we are interested in understanding the impact of generalized association rules in classification processes. For that purpose, we investigate the use of generalized association rules,i.e., rules in which the conclusion is a disjunction of attributes. We propose a method which directly mines nonredundant generalized association rules, possibly with exceptions, by using the recent developments in condensed representations of pattern mining and hypergraph transversals computing. Then we study the impact of using such rules instead of classical ones for classification purposes. To that aim, we view generalized rules as rules with negations in the premise and possibly concluding on a negative class attribute. To study the impact of such rules, we feed the standard CMAR method with these rules and we compare the results with the use of classical ones.

Domaines: Intelligence artificielle [cs.AI]; Modélisation et simulation; Système multi-agents [cs.MA]

François Rioult, Bruno Zanuttini, Bruno Crémilleux. Nonredundant Generalized Rules and Their Impact in Classification. Advances in Intelligent Information Systems, Springer, pp.3-25, 2010. ⟨hal-00944355⟩