Individual Preference Learning with Collaborative Learning Framework

Thesis (Ph.D.)--University of Washington, 2020Smart, personal devices that interact with individuals make it possible to trigger desired behavioral changes with personalized incentives. Personalized incentives are the incentives that suit an individual's preferences. In this dissertation, individual preferences refer to a set of parameters describing how the individual values each influential factor in a travel alternative. To trigger behavioral changes with personalized incentives, a model that can accurately and efficiently estimate an individual's preferences from his behavior data is required. Two challenges exist in individual preference learning. For the first, the number of observations available from each individual for individual preference learning is limited. This issue causes difficulties in preference updating. For the second, the observability of the choices made is limited. This is because that it is not possible to directly observe the preference parameters -- the only information that can be observed is an individual's choice-making behavior. The two challenges prevent the use of traditional preference-learning techniques such as advanced econometric models (e.g., discrete choice models) derived from Random Utility Maximization (RUM).Other techniques such as machine learning also cannot be applied for similar reasons. New methods are needed for individual preference learning. This dissertation contributes to the existing literature in travel behavior studies by proposing individual preference learning methods such that personalized incentives could be accurately estimated to trigger behavioral changes, and proposing a design of an online experiment to collect travel behavior data. Specifically, two research questions are of interest: (1) What methodology could be used to learn an individual's preferences with only a few observations of choices made by him? (2) How to collect individuals' choice data to test the method proposed in the dissertation in terms of triggering individual behavioral changes with personalized incentives? In the dissertation, the behavior data is collected via a carefully designed online experiment utilizing the AMT (Amazon Mechanical Turk) platform. Considering the validity and reliability of the data, the dissertation contributes to the travel behavioral study in: (1) a full factorial design of a randomized experiment with two factors (commuting time and work flexibility, each with three levels) utilizing the online platform of AMT (Amazon Mechanical Turk) to collect individuals' travel choices on departure time in a sequence of hypothetical scenarios, and (2) a design of data quality control strategies, which refers to the design of some methods to reduce and identify the low-quality data collected in the experiment. These data quality control methods, such as understanding check, response consistency check, responding time record, and social desirability scale, can be applied to other online experiments and behavioral studies. To learn an individual's preference from a few choices made by him, a model structure that integrates a time-varying model and the collaborative learning model is proposed in the dissertation. The time-varying model is used to replace the original constant preference parameter to a time-dependent function, allowing an individual's preferences to fluctuate in his choice-making process. The collaborative learning model can exploit the underlying canonical structure of individuals' preference variation in a heterogeneous population. Specifically, the collaborative learning model could identify several patterns of preference changes (known as "canonical models") that exist in the population. With the canonical models, each individual's preference change can be expressed by a linear combination of all those canonical models. Considering the model's computation time, an online updating strategy for the proposed model is also proposed, such that individual preferences could be learned accurately and efficiently. Detailed specifications of two different formulations of the time-varying model are presented in the dissertation, with some explorations on model properties with simulations. The models are also applied to the real-world dataset collected in the online experiment. Results show that the proposed models can achieve higher accuracy in parameter learning and behavioral prediction than traditional preference learning models such as the logit model and the mixed logit model