NEB-Route: A Multi-Modal Framework for Missing Persons Search
First author · Submitted, under revision (2026)
Bayesian sequential search framework for locating missing persons in urban environments, combining personalized Hawkes processes, graph neural networks (GCN), and behavioral homophily. Reduces average detection time by 19.4% over the state of the art (DeepMove-Seq) on the HuMob dataset (Tokyo, 100,000 individuals).
Details
- Problem: missing-persons search in urban areas often succeeds less than 5% of the time within the first hours, despite abundant mobility data generated by smart cities.
- Reformulation: localization is framed as dynamic investigation-resource allocation under constraints, via sequential Bayesian decision-making that explicitly incorporates negative observations (non-detections) to reallocate search attention.
- Three combined information sources: individual routines (personalizable Hawkes processes), collective flows (graph neural networks / GCN), and behavioral homophily (socio-spatial similarity between individuals).
- Evaluation: 100 simulated missing-person scenarios on the public HuMob dataset (100,000 individuals, Tokyo, 23.45 million location points).
- Results: 19.4% reduction in average detection time versus the state of the art (DeepMove-Seq), 58% detection probability within the first 10 iterations (vs. 46%).
- Ablation study: personalization of the Hawkes model is the most critical component (+15.5% detection time if removed), followed by behavioral homophily (+10.3%).
- Acknowledged limitations: behavioral-regularity assumption, dependency on mobility history, ethical questions around large-scale mobility data use.