The proliferation of passive mobility big data sources, including mobile phones, smart cards, GPS trajectories, social media and other sensing technologies, has transformed how mobility patterns are observed. These sources generate massive amounts of spatiotemporal information on a continuous, large scale, revealing when and where individuals travel but not trip purpose or the type of activities undertaken at each stay. This necessitates robust activity/trip purpose inference methods to bridge this gap. We provide a critical review of existing inference methods for heterogeneous big data sources based on a systematic review of the literature. These methods are grouped into six independent categories: Rule-based or heuristic methods, Bayesian spatial interaction, Supervised, Unsupervised, Semi-supervised, Reinforcement learning models. For each category, we critically examine the basic assumptions, strengths, limitations, data applicability, and contextual considerations. Also, we assess hybrid strategies and options that combine some of the six in an inference task. This synthesis indicates that mobile phone, smart card, and GPS data dominate current research, while opportunities exist to integrate under-utilised sources like Wi-Fi sensing and crowdsourced social media. Primary activities (e.g., home, work/study) are relatively straightforward to infer, but irregular or context-dependent activities remain challenging. Future avenues include expanding the use of semi-supervised and reinforcement learning, developing innovative hybrid frameworks, and incorporating richer contextual and textual information. We conclude with a practical roadmap for aligning method selection with data characteristics, research objectives, and decision criteria to guide planners, policy-makers and practitioners in the selection and implementation of accurate, scalable, and context-aware inference methods.
Abdi et al. (Fri,) studied this question.