Traveler time allocation between activities and travel: A computational graph-infused machine learning framework with bi-mode MDCEV model

Multiple discrete–continuous extreme value (MDCEV) model has received increasing attention in modeling travelers’ time allocation. However, the MDCEV model primarily focuses on activity choices that generate increased utility over time, overlooking the fact that travel, as a derived activity, is typically something people prefer to save time on. To bridge the research gap, we introduce the bi-mode utility multiple discrete–continuous extreme value (BMU-MDCEV) model. This model seamlessly integrates a diminishing function for travel with the classical utility function, underscoring the principle that individuals typically aim to minimize travel time. Inspired by the fact that both choice modeling and machine learning (ML) involve non-convex optimization processes, this study implements an ML-based computational graph (CG) mechanism to provide reliable and efficient parameter estimates for the proposed model. This approach emphasizes the integration of theory- and data-driven methods within the context of multiple discrete–continuous choices. Validated using the National Household Travel Survey (NHTS) 2017 dataset, the CG-enhanced BMU-MDCEV model effectively uncovers socioeconomic heterogeneity and captures the substitution and complementarity in time allocation patterns. Our analysis of the marginal utility of various travel types reveals a positive correlation between travel time tolerance and activity satiation for discretionary activities (e.g., leisure, shopping). Conversely, individuals tend to reduce unnecessary travel time for mandatory daily activities (e.g., home-based activity, work), regardless of the degree of activity satiation. By shedding light on the nuanced mechanisms behind individuals’ time allocation, our proposed method paves the way for informed transport management strategies that respond more effectively to individual behavior.