Learning Query Rules for Optimizing Databases with Update Rules

In this paper, we consider an approach to updating Datalog neg databases containing two kinds of rules, namely query-rules and update-rules that can be seen as constraints. Update-rules are used to compute the side eeects of updates, and these side eeects act as exceptions to derivations through query-rules. Additionally, every fact to be inserted or to be deleted is stored in the database. In this way, our approach handles insertions and deletions over intensional predicates in a sound and deterministic way. However, two important problems occur: rst, the overhead incurred by the storage of inserted and deleted facts may be important, and, second, the intensional database (i.e., the query-rules) may not \correspond" to the extensional database (i.e., the facts stored in the database together with the update-rules). In order to cope with these diiculties, we propose to use Machine Learning techniques in order to compute new query-rules, starting from the semantics of the database. Following our approach, true facts and false facts of the semantics of the original database are given as input to the system which returns a new database containing fewer exceptions and whose semantics contains the semantics of the original database. Dans cet article, nous consid erons une approche de mise a jour des bases de donn ees de type Datalog neg contenant deux types de r egles : des r egles dites de requ^ ete et des r egles dites de mise a jour, ces derni eres pouvant ^ etre vues comme des contraintes. Les r egles de mise a jour sont utilis ees pour calculer des eeets de bord engendr es par les mises a jour, peut ne plus \correspondre" a la base extensionnelle (i.e., les faits stock es et les r egles de mise a jour). AAn de r esoudre ces diicult es, nous proposons d'utiliser des techniques issues de l'Apprentissage Symbolique Automatique aan d'engendrer de nouvelles r egles de requ^ ete a partir de la s emantique de la base de donn ees. Selon notre approche, les litt eraux de la s emantique de la base de donn ees sont fournis au syst eme qui retourne une nouvelle base de donn ees contenant moins d'exceptions et dont la s emantique contient la s emantique de la base originelle. Mots-Cl es: Apprentissage symbolique automatique, base de donn ees d eductive, pro-grammation logique inductive, mise a jour, s emantique bien fond ee.