Réseaux GAI pour la prise de décision

This paper deals with preference elicitation and preference-based recommendation (choice and ranking) in the context of multiattribute utility theory. We focus on the generalized additive decomposable utility model (GAI) which allows interactions between attributes while preserving some decomposability. We first present a systematic elicitation procedure for such utility functions. This procedure relies on a graphical model called a GAI-network which is used to represent and manage independences between attributes. Then, we propose a choice procedure as well as a ranking procedure relying on a GAI-network to solve efficiently both optimization and ranking problems over a product set. We show that GAI networks can jointly integrate constraints and preferences and can thus be efficiently exploited in this case to compute constrained choices and ranking. Finally, we provide results of numerical experiments to highlight the practical efficiency of our approach. MOTS-CLÉS : réseaux GAI ; modèles graphiques; élicitation de préférences; rangement.