Learning Modulo Theories for constructive preference elicitation

This paper introduces CLEO, a novel preference elicitation algorithm capable of recommending complex configurable objects characterized by both discrete and continuous attributes constraints defined over them. While existing techniques focus on searching for the best instance in database candidates, CLEO takes constructive approach to recommendation through interactive optimization space feasible configurations. The assumes minimal initial information, i.e., set catalog attributes, defines decisional features as logic formulae combining Boolean algebraic attributes. (unknown) utility decision maker is modeled weighted combination features. iteratively alternates step, where pairs candidate configurations are selected based current model, refinement step refined incorporating feedback received. leverages Max-SMT solver return optimal according model. implemented learning rank, sparsifying norm used favor selection few informative combinatorial A major feature that it can recommend hybrid domains (i.e., including numeric attributes), thanks use technology, while retaining uncertainty decision-maker's noisy feedback. In so doing, adapts recently introduced modulo theory framework setting. formulation function coupled with capabilities 1-norm regularization allow effectively deal DM high expressiveness. Experimental results tasks show ability quickly identify configurations, well its capacity recover from suboptimal choices. Our empirical evaluation highlights how outperforms state-of-the-art Bayesian when applied purely task