Keywords

1 Introduction

On March 11, 2011, a massive “9.0" earthquake struck the Pacific Ocean, approximately 130 km off Sendai City, Japan. This event serves as a case study for our research [8, 12, 13]. We aim to jointly model the historical visitation patterns of functional areas and decision-making behavior following such a significant earthquake. Numerous studies have explored mobility patterns using historical visit data collected through smartphones. These studies have been instrumental in analyzing people’s behavior during disasters [2, 3, 17, 18]. For instance, [19] developed a probabilistic model to simulate population evacuation across complex geographic features in Japan in response to future disasters. In addition, [3] demonstrated that location data can be utilized to track the distribution of earthquake risk areas, people’s emergency responses, and overall behavioral patterns during disasters. However, previous works often overlooked the significance of functional regions [23] in understanding human mobility during disasters. These functional regions contain crucial information that can enhance our comprehension of human behavior in such scenarios. In conclusion, our study seeks to bridge this gap and includes an empirical prediction method to estimate how individuals chose to return home following the Tohoku earthquake. By jointly considering historical visitation data of functional areas and decision-making behavior, we aim to gain deeper insights into human mobility patterns after a major earthquake event.

More than a decade ago, seldom work focused on analyzing the human mobility pattern in natural disasters. Fortunately, mobile sensing technologies have been widely applied in various fields. This resulting multiple works [4, 21] have been proposed to predict human mobility in large-scale disasters (such as earthquakes, tsunamis, and hurricanes). However, they may lack some necessary to analyze a human mobility strategy. Since large-scale disaster is rare, it’s a challenge to understand human mobility from such data. Although there are a lot of related works for analyzing [15, 22] after a big earthquake, few pieces of research consider the influence of personal factors on decision-making. In addition, the knowledge of what key factors impel a person to choose a returning home strategy is important for analyzing the human decision after a big earthquake.

Here, we present an overview of our framework. The system comprises two main steps: data mining and explainable knowledge.

In the data mining step, we begin by constructing a staying cuboid that collects the number of staying points generated during returning home trajectories for each region. And we could obtain different region topics of those historical personal visits by analyzing surrounding POI distributions. Regarding the transition mode, we employ a map-matching algorithm [16] for each trajectory to map the GPS points onto road segments. This allows us to determine the transportation modes based on features within each segment, such as railways, expressways, and footpaths.

After gathering the necessary data, we define the return home mode and departure time from the workplace as a multitask objective function. This approach helps us generate an explainable result to support our analysis effectively. Finally, we present the feature importance histogram to demonstrate the effectiveness of our proposed framework.

2 Related Work

The topics of human behavior prediction during disaster (e.g., crowd panics [14], fires [6, 9, 10], floods [5]) have received numinous attention in recently with a focus on small-scale or short-term emergencies. However, research on the dynamics of population movements across the country in large-scale disasters (such as earthquakes, tsunamis, and hurricanes) is minimal [11]. Meanwhile, the extensive volume of work has brought about a significant shift in the mobility pattern [2, 3, 17, 18] derived from historical data collected by smartphones. This wealth of data is now being harnessed to analyze human behavior during stages of disasters. One noteworthy application is demonstrated by [19], who devised a comprehensive probabilistic model to simulate population evacuation across intricate geographic terrains in Japan, foreseeing potential future calamities. Moreover, by leveraging location data, [3] discovered that it is feasible to monitor the distribution of earthquake risk areas, assess people’s emergency responses, and track their behavior during such critical events. Reference [20] points out that although people now will go further and faster than before, most of their time in people’s lives will still locate in important places, such as home, workplace. However, those work ignore the functional region [23] in practice, which may lack some necessary information for understanding human mobility. For example, people’s emergency measures are closely related to their located region, and historical visiting.

3 Data, Tasks and Feature Importance Analysis

3.1 Data Descriptions

POI Data. In our study, we gathered the Telepoint Pack DB of POI data in February 2011, which was provided by ZENRIN DataCom Co., Ltd [1]. The original database consists of records containing registered landline telephone numbers along with their associated coordinates (latitude, longitude) and industry category information. For the purpose of our research, we considered each “telepoint" as a specific point of interest (POI). To facilitate our analysis, we categorized the POIs into five broad types (i.e., commuting, public place, shopping, restaurant, and entertainment).

Human Mobility Data. In our study, we gathered a large GPS log dataset anonymously from approximately 1.6 million real mobile phone users in Japan. The data was collected over a 12-day period, specifically from March 1st to March 12th, 2011. The data collection was conducted by two entities: the mobile operator NTT DoCoMo, Inc. and the private workplace ZENRIN DataCom Co., Ltd. The mobile phone users provided their consent for the data collection process. To ensure privacy protection, the collected data were processed collectively and statistically. This procedure ensured that sensitive information such as gender or age was concealed and not accessible for analysis. The positioning function on the users’ mobile phones was activated every 5 min by default. However, data acquisition could be affected by factors like signal loss or low battery power. When a mobile phone user remained stationary at a location, the positioning function was automatically turned off to conserve power. It’s important to note that the dataset’s age distribution slightly favors young users, as they tend to prefer mobile phones with positioning functionality compared to other age groups (e.g., the elderly). The representativeness of our dataset was verified through previous work, where its quality was evaluated [7]. For our specific analysis of human mobility after the Tohoku Earthquake, we focused on the period from March 1st to March 11th, 2011. This period allowed us to capture and study the mobility patterns and behaviors of individuals following the earthquake event.

Fig. 1.
figure 1

Traffic volume and ratio changes of each travel mode choice (“Stay", “Walk", “Bike", “Train" and “Car") before/during/after Tohoku Earthquake (March 10th/11th/12th) in Great Tokyo area. The earthquake happened at 14:46 PM on March 11th, and the horizontal axis is the 24 h of one day.

Full size image

3.2 Returning Home Strategy-Based Tasks

In this study, for those mobile phone users who returned home from their companies after the Tohoku earthquake, we divide proposed decision-making strategies into two main tasks. The first is the travel mode choice prediction, and then is the schedule prediction of departure time from the workplace. The travel mode and departure time from the workplace comprise the returning home strategy.

First, we conduct two sub-tasks regarding the travel mode choice prediction. As is shown as Fig. 1, most subway lines were halted due to the damage to the big earthquake. This caused the train-based travel mode to decline sharply. The primary railway transportation network would recover until 11:00 PM on March 12th. This raises a question: when it comes to the big earthquake if one mobile phone user often came back home from the workplace by railway before, how did him/her make choices for returning home on March 11th? The possible choice for him/her would be Choice 1 - Walk directly; Choice 2 - Wait for the recovery of the train and continue to choose the railway and Choice 3 - Not go home to a hotel/refuge. As a result, sub-task 1 is a three-class classification problem. Except for the decision-making strategy prediction for those who often come back home from their companies by railway, we also conduct a three-class travel mode prediction for all mobile phone users directly, i.e., train-based, walk-based (including Walk and Bike), car-based travel mode.

Secondly, estimating departure times from workplaces is a crucial aspect of comprehending decision-making behaviors during significant disasters. For instance, during the Tohoku earthquake that struck at 14:46 (Japanese standard time), numerous companies allowed their employees to leave work immediately. Nevertheless, with the transportation network severely disrupted, a considerable number of individuals opted to remain at their workplaces, anticipating the restoration of public transportation. As a result, we undertake the task of estimating departure times from workplaces for these mobile phone users.

Fig. 2.
figure 2

Feature importance analysis for travel mode choice inference. Here, we implement the feature importance analysis of sub-task 1.

Full size image

3.3 Feature Importance Analysis

Although we have already known the feature importance of “HV" from the inference results, we also provide some interesting and reasonable discoveries for decision-making strategies based on feature importance analysis. In Fig. 2 and Fig. 3, we show the all considered factors/features importance values for two inference tasks, respectively. These values were captured by the Shapash library when using all features in LightGBM model. The higher values of the figures mean the big feature importance. At first, for the travel mode choice inference (sub-task 1), users’ locational information (especially “distance") has relatively higher important values than users’ historical visit information. But if one user intended to choose to wait for the recovery of the train, the feature “Topic_restaurant" has highly important values. It’s reasonable because phone users who often visited the restaurant before the earthquake were more likely to choose to wait for the recovery of the train in nearby bars/restaurants. On the contrary, if one user chooses to walk directly, the feature “Topic_restaurant" should also be important. Our feature importance results for class of “walk" in Fig. 2 validates this assumption. Especially, the class of “Not go home" only has the highest importance values for feature “his_stay", which represents the total stay time of historical visits, but this class is not very related with the topics of historical visits (relatively low importance values with “Topic_xxx"). Secondly, for the departure time estimation from workplace, the feature importance results also represent the same pattern with the class of “Not go home" in travel mode inference, which equally shows the high importance values in “his_stay" when considering the “HV" factors. In sum, this suggests that the decisions of mobile phone users were easily influenced by “his_stay". But the topics of historical visits only show the feasibility in specific travel mode choice. For example, “Topic_Restaurant" only shows strong feature importance in terms of the classes of both “Train" and “Not go home" under travel mode choice inference (Fig. 3).

Fig. 3.
figure 3

Feature importance analysis for departure time estimation from workplace.

Full size image
Fig. 4.
figure 4

The comparative feature importance analysis for total mobile phone users and those phone users whose work location were nearby Shinjuku station. Here, the analyzed task for feature importance is the departure time estimation from workplace.

Full size image

4 Conclusion

In this study, we collected large-scale GPS log datasets and analyzed returning home strategy of mobile phone users during Tohoku earthquake. Our work emphasized the importance of understanding the topics of historical visit, to obtain a holistic view regarding the human behaviors of different topics/types. Then all influence factors from users’ historical visit, historical travel model choice and locational features could be informative about the decision-making strategies inference during disasters. We define and evaluate two main prediction tasks (i.e., travel mode choice prediction and departure time estimation from company) for returning home strategy during disasters, showing the feasibility of using smartphone sensing to detect historical visits. In addition, we conduct an explainable analysis about the key factors that impel phone users to make decisions. We believe these findings could be useful for ubicomp and the government towards implementing future mobile disaster evacuation systems by effectively understanding human decisions during big disasters.