The Role of Model Selection in Preference Learning

Abstract

Learning preferences of an agent requires choosing which preference representation to use. This formalism should be expressive enough to capture a significant part of the agent's preferences. Selecting the right formalism is generally not easy, as we have limited access to the way the agent makes her choices. It is then important to understand how ``universal" particular preference representation formalisms are, that is, whether they can perform well in learning preferences of agents with a broad spectrum of preference orders. In this paper, we consider several preference representation formalisms from this perspective: lexicographic preference models, preference formulas, sets of (ranked) preference formulas, and neural networks. We find that the latter two show a good potential as general preference representation formalisms. We show that this holds true when learning preferences of a single agent but also when learning models to represent consensus preferences of a group of agents.