Comparing four different preference modeling

Multicriteria decision support systems are often based on the use of a particular set function which values the influence on the decision of not just one criterion, but also a coalition of criteria. From a training set used to find an appropriate set function, it is possible to mimic the decision-making behaviour of one or more experts. However, it can be more reliable to ask an expert to express a preference between two training examples rather than an overall score for each example, particularly when the examples arrive in a sequential stream. In this paper, we propose to model the preference relation by a game and to train it from a sequential stream of examples. Four different games are studied : a general game, a 2-additive game, a linear game, and a maxitive game. Experiments are conducted with common datasets for regression methods.