Affiliations: Diagnogene, Aurillac, France, LIMOS, UMR 6158 CNRS,
Univ. Clermont-Ferrand II, France. E-mail: agier@isima.fr | LIRIS, UMR 5205 CNRS, INSA Lyon, France.
E-mail: jean-marc.petit@insa-lyon.fr | Graduate School of Information Science and Electrical
Engineering, Kyushu University, Japan. E-mail: suzuki@i.kyushu-u.ac.jp
Abstract: The notion of rules is very popular and appears in different
flavors, for example as association rules in data mining or as functional
dependencies in databases. Their syntax is the same but their semantics widely
differs. In the context of gene expression data mining, we introduce three
typical examples of rule semantics and for each one, we point out that
Armstrong's axioms are sound and complete. In this setting, we
propose a unifying framework in which any
"well-formed" semantics for rules may be integrated.
We do not focus on the underlying data mining problems posed by the discovery
of rules, rather we prefer to discuss the expressiveness of our contribution in
a particular application domain: the understanding of gene regulatory networks
from gene expression data. The key idea is that biologists have the opportunity
to choose – among some predefined semantics – or to define the meaning of their
rules which best fits into their requirements. Our proposition has been
implemented and integrated into an existing open-source system named MeV of the
TIGR environment devoted to microarray data interpretation. An application has
been performed on expression profiles of a sub-sample of genes from breast
cancer tumors.
Keywords: Rules, implications, Armstrong's axiom system, data mining
