Services objectivization: A ranking approach

Objectivization is a crucial task arising in a wide variety of industrial fields that aims at optimizing a certain service in terms of "customer satisfaction". In this paper focus is on car drivability objectivization and more precisely on the problem of calibrating the acceleration feeling. Our approach is based on statistical ranking of a sample of vehicles characterized by a series of physical parameters, so that cars whose drivability is positively evaluated ideally appear at the top of the list. A novel ranking method is used for this purpose, called TreeRank, that may be viewed as a recursive implementation of a cost- sensitive version of the celebrated classification algorithm Cart. When applied to a data sample made of pairs (X, Y) where Y is a binary label indicating subjective evaluation of a car's drivability and X its characteristics, this specific nonparametric partitioning technique not only outperforms standard methods based on statistical modelling of the posterior distribution but also yields easy-to-interpret models. Additionally, we show how to apply bootstrap aggregating techniques in this context in order to enhance ranking accuracy, the performance of the resulting model comparing favorably to a currently used method based on the Lasso procedure.