Abstract
In the wake of natural disasters, large-scale population movements present significant challenges for social stability and disaster management. This study analyzes human mobility strategies following the 2011 Tohoku earthquake in Japan using a massive dataset of smartphone location logs from 1.6 million users. By integrating Point of Interest (POI) distributions and historical visitation patterns, we developed a multitask learning model to predict individual "returning home" strategies—specifically, the choice between walking, waiting for public transit, or staying in place. Our findings highlight that historical behavior and regional functional "topics" are critical predictors of human decision-making under extreme stress.
Introduction
When the magnitude 9.0 Tohoku earthquake struck, transportation networks across Japan collapsed, leaving millions to decide whether to attempt a journey home on foot, wait for infrastructure recovery, or seek local shelter. Traditional disaster research often lacks the granular, individual-level data necessary to understand the personal factors driving these high-stakes decisions. This paper addresses this gap by leveraging "big mobility data" to create explainable models of human decision-making during catastrophic events.
Methodology
The research framework combines data mining with explainable AI (XAI) through several key steps:
- Region Topic Modeling: Geographic grids were categorized into five functional "topics" (commuting, public places, shopping, restaurants, and entertainment) based on POI distributions.
- Staying Cuboids: Historical visitation (HV) patterns and stay points were aggregated for each user to capture their typical behavior in various regions.
- Map-Matching: GPS pings were aligned with physical infrastructure (railways, footpaths) to accurately identify transportation modes.
- Multitask Learning: A LightGBM-based model was trained to simultaneously predict travel mode choice (Walk, Wait, or Stay) and departure time from the workplace.
- Explainability: The Shapash library was used to identify feature importance, revealing the underlying logic of the model's predictions.
Mobility Shifts and Results
The impact of the earthquake on daily routines was immediate and profound. Analysis of the GPS logs reveals a sharp pivot in transportation modes as the rail network became unavailable.
As shown in the figure above, the "During" phase (March 11th) shows a massive spike in "STAY" and "WALK" behaviors compared to the pre-disaster baseline. By March 12th, while some normalcy began to return, the distribution of mobility modes remained significantly altered.
Key Predictors of Strategy
- The "Restaurant" Affinity: One of the most interesting findings was that users who frequently visited restaurants or bars prior to the quake were more likely to wait for train recovery rather than walk. This suggests a preference for waiting in familiar, service-rich environments.
- Historical Visitation (HV): While distance is the primary factor for those choosing to walk, historical stay time (his_stay) is the dominant predictor for those who choose not to return home at all. If a person has a high "stay" value in their workplace region, they are statistically more likely to remain there or in a nearby shelter.
- Infrastructure Sensitivity: The model successfully captured the transition from rail to walking or staying as a direct response to infrastructure failure following the 14:46 PM quake.
Conclusion
This study demonstrates that smartphone sensing data, combined with historical behavioral analysis, can effectively model complex human decision-making during disasters. By understanding the "topics" of regions and individual visitation histories, government agencies can better predict crowd behavior and optimize the placement of emergency shelters and resources. Future work will aim to address demographic biases in early smartphone data to create even more inclusive disaster response models.
Summary updated