An online updating method for time-varying preference learning

The rapid proliferation of smart, personal technologies has given birth to smart Transportation Demand Management (TDM) systems that can give personalized incentives to users. This personalization capacity builds on accurate modeling of user behaviors; however, in practice, a user’s behavior data is often limited, and his preferences in the discrete choice-making process may change or evolve. In this paper, we propose a new online-updating model that can accurately and efficiently estimate an individual’s preferences from his discrete choices. Our model is built on the concept of canonical structure, where a set of canonical models are identified as the common preference patterns shared by the whole population, and a membership vector is also identified for each individual to capture the degrees of the resemblance of his preferences to those common preference patterns. To allow preference to vary in the choice-making process, a time-varying model can be integrated with the canonical structure. In the current study, we use a simple cubic polynomial model with a single variant and show the detailed formulation of the integrated model. An online-updating strategy is also proposed, such that it is possible to update the parameters partially in practice. The proposed model is suitable for modeling a heterogeneous population with insufficient data from each individual. Both simulation studies and a real-world application are taken in the current study. The results show that comparing with other frequently used models, the model we proposed has the highest accuracy in preference learning and behavior prediction. •We develop a model to learn an individual’s preference in sequential choices.•The proposed approach is able to capture preference changes.•The proposed approach permits learning from limited data.•An online updating strategy is proposed to effectively update the estimates.