A feature-based framework for evaluating synthetic human mobility

Generating realistic human mobility trajectories is essential for applications in urban analytics, transportation planning, and privacy-preserving data sharing. Evaluating the quality of synthetic data remains challenging. This study introduces a feature-based evaluation framework that characterizes trajectories through a unified set of statistical, geometric, and temporal descriptors. The framework is applied to benchmark GAN- and diffusion-based generative models using three real-world urban datasets with distinct spatial structures. Region-specific fine-tuning enhances realism, while persistent discrepancies in multi-scale entropy coefficients reveal challenges in modeling transitions between dwell and trip states. Incorporating road network information after generation provides limited benefit, suggesting that spatial constraints should be embedded during training. These findings highlight the influence of trajectory length, data quality, and explicit state modeling on generative performance. The study establishes a transparent feature-based approach connecting generative modeling and mobility analysis, supporting the creation of synthetic agents for data-driven urban design and policy evaluation.

Author dc:creator

• Han, Jin

Advisor dc:contributor.supervisor

• Stanley, Kevin

• Urban mobility
• trajectory simulation
• Evaluation
• Spatial behavior

Statement dc:rights

• Available to the World Wide Web

Language dc:language.iso

en, English

Handle dc:identifier.uri

https://hdl.handle.net/1828/23567

OAI identifier oai:identifier

oai:dspace.library.uvic.ca:1828/23567