Stacked Hybrid Discrete Choice Models for Airline Itinerary Choice

This study develops a methodology to train and apply a hybrid stacked discrete choice model for airline itinerary choice. This stacked model framework includes a data-driven component (i.e., gradient boosting machines) as well as a theory-driven component (i.e., utility maximization using generalized extreme value models). The resulting ensemble model combines attractive features of each, including the ability to conform to complex nonlinear relationships among itinerary characteristics, as well as the ability to leverage an analyst's understanding of travel behavior tendencies and the natural relationship among itineraries. Using a real industry dataset containing purchase information for approximately 10 million air travelers, it is demonstrated that the resulting model outperforms either the gradient boosting or utility maximization modeling paradigm alone in forecasting air traveler choice behavior. Implementation of this model can be achieved using efficient open source tools including XGBoost and Larch, and requires relatively modest additional effort by an analyst above and beyond the effort to use either tool alone.