Behaviorally Informed Machine Learning for Human Mobility

Abstract

Large-scale digital trace datasets hold considerable promise for long-range transportation planning, offering the potential to observe mobility at metropolitan scales with far greater temporal and spatial resolution than traditional household travel surveys while capturing the regularities in daily travel behavior that underlie trip-making. Passively-collected mobile (PCM) data (e.g., location signals from smartphones and in-vehicle GPS) are central to this promise. Their usefulness, however, is limited by discontinuities in individual trajectories, privacy constraints that restrict data sharing and integration, and representativeness biases that distort inferred patterns of regional travel demand and travel behavior across population groups. This dissertation addresses these limitations through four contributions. First, it develops a multi-task Gaussian Process-based imputation method (grounded in recurring daily, weekly, and seasonal travel behavior patterns) capable of handling both short- and long-duration gaps in GPS traces, significantly improving the completeness and usability of mobility data. Second, it introduces an individualized, physics-regularized learning framework that produces high-fidelity mobility traces reflective of observed movement patterns. These generated trajectories can be scaled to build richer, more diverse mobility datasets for developing and validating activity-based models. Third, it investigates the predictive signal linking mobility patterns as expressions of travel behavior to sociodemographic attributes that shape those behaviors, and develops imputation strategies for enriching PCM datasets with these inferred labels. This enrichment supports both more detailed planning analyses and a clearer diagnosis of representativeness biases in passively collected data. Finally, through a qualitative study of long-range transportation planners, this dissertation investigates barriers to the adoption of big data products and provides recommendations for their effective integration into planning processes. Together, these contributions bridge methodological advances in machine learning with insights from travel behavior research and the practical needs of public agencies, offering a more transparent and behaviorally coherent foundation for data-driven planning and travel behavior analysis.