Modeling Bias in Passively Collected Mobility Data: A Data Generation Process Perspective

Abstract

Big data, particularly passively collected mobile data from smartphones, is increasingly used in transportation research to infer human mobility patterns and evaluate system performance. Despite its broad coverage and high temporal resolution, such data are generated through opaque and non-probabilistic processes, introducing systematic biases and missingness that fundamentally constrain downstream analysis. Rather than treating bias as a single measurable quantity, this dissertation conceptualizes bias as a property of the data generation process (DGP) that governs what information is observable, recoverable, and generable from passively collected data. Using smartphone app-based mobility data as a case study, this research proposes a two-stage DGP framework consisting of a population identification stage followed by an observation stage. The first stage captures demographic selection bias arising from heterogeneous smartphone ownership and differential identification likelihood across demographic groups, while individual-level demographic attributes remain unobserved due to privacy constraints. The second stage induces temporal and spatial missingness in individual mobility records, where true mobility trajectories are not directly observable. These mechanisms are formalized to clarify how bias arises prior to any modeling or inference. Within this framework, the dissertation examines the extent to which latent mobility structure can be recovered under biased and incomplete observations. Bayesian inference is used to estimate demographic composition under anonymity constraints, and recommendation system-based matrix factorization methods are evaluated for imputing missing spatiotemporal observations using simulated ground truth. The results indicate that such inference methods provide limited and context-dependent recovery, highlighting structural limits imposed by the data generation process. The dissertation further investigates whether generative models can serve as alternatives to simulation-based data generation. The dissertation further explores generative models as an alternative to simulation-based data generation across multiple mobility representations, including GPS trajectories, zone-level trajectories, and trip-level data. The findings show that generative performance strongly depends on the level of data abstraction, with trip-level representations being substantially more learnable than trajectory-level data, and that commonly used baseline evaluation metrics may be insufficient for capturing spatiotemporal structure. Collectively, this dissertation demonstrates that bias in passively collected mobility data is understood as a constraint on inference and generation rather than a defect that can be universally mitigated. The results underscore the importance of explicitly modeling data generation processes and representation choices when using passive data for transportation analysis and planning.