Abstract
Predicting the evacuation decisions of individuals before the disaster strikes is crucial for planning first response strategies. In addition to the studies on post-disaster analysis of evacuation behavior, there are various works that attempt to predict the evacuation decisions beforehand. Most of these predictive methods, however, require real time location data for calibration, which are becoming much harder to obtain due to the rising privacy concerns. Meanwhile, web search queries of anonymous users have been collected by web companies. Although such data raise less privacy concerns, they have been under-utilized for various applications. In this study, we investigate whether web search data observed prior to the disaster can be used to predict the evacuation decisions. More specifically, we utilize a "session-based query encoder" that learns the representations of each user's web search behavior prior to evacuation. Our proposed approach is empirically tested using web search data collected from users affected by a major flood in Japan. Results are validated using location data collected from mobile phones of the same set of users as ground truth. We show that evacuation decisions can be accurately predicted (84%) using only the users' pre-disaster web search data as input. This study proposes an alternative method for evacuation prediction that does not require highly sensitive location data, which can assist local governments to prepare effective first response strategies.
Introduction
Effective disaster response—such as the strategic allocation of emergency supplies or the timely opening of shelters—depends heavily on the ability to anticipate who will actually evacuate. Traditionally, researchers have relied on post-disaster surveys or real-time mobile phone GPS data to model and predict this behavior.
However, the use of GPS data presents significant hurdles, primarily regarding privacy. Real-time location tracking is increasingly restricted and controversial. Furthermore, while location data indicates where an individual is, it does not necessarily reveal their intent. This research explores whether web search queries, which are less privacy-sensitive and directly reflect a user's information-seeking process, can serve as a viable proxy for predicting high-stakes mobility decisions before a disaster alert is even issued.
Methodology
The researchers proposed a framework that transforms raw search queries into "user representations"—mathematical vectors that capture search intent. The process is divided into three primary stages: training a query encoder, obtaining user representations, and classification.
Session-based Query Encoding (SQE)
The core of the approach is a character-based 3-layer Long Short-Term Memory (LSTM) network. This model was trained on a massive dataset of 301 million search sessions to predict the "next query" in a sequence. This training forces the model to learn the latent semantics of search behavior, such as understanding that searches for "sandbags" and "river level" are both related to flood preparation.
To handle multiple queries from a single user, the team employed a "Session-based Multiple Query Encoder" (SMQE). This hierarchical architecture feeds the output of individual query encodings into a second LSTM layer to capture the context of an entire search session. Feature selection was optimized by comparing the most recent queries against "high-importance" queries identified via tf-idf.
Datasets and Validation
The study utilized a large-scale dataset from Yahoo Japan Corporation, containing 804 million queries collected between 2015 and 2018. Mobile phone GPS data was used strictly as ground truth for validation. By applying anomaly detection to GPS logs, the researchers determined if a user actually evacuated (moved to an unusual location like a shelter) during the devastating floods of July 2018 in the Okayama and Hiroshima prefectures.
Results and Discussion
The analysis demonstrated that search behavior is a powerful leading indicator of evacuation. Individuals began searching for disaster-related information, such as traffic conditions and flood levels, well before official evacuation alerts were issued.
The SQE model exhibited a sophisticated understanding of latent semantics. It could distinguish between homonyms and recognized that searches for seemingly unrelated terms, like "all-night restaurants," were often proxies for seeking temporary shelter. Furthermore, the model picked up on specific, sparsely searched local landmarks that traditional keyword-matching methods would likely ignore.
The best-performing model achieved a predictive accuracy of 84% and an AUC of 0.837. This represents a 10% improvement over standard one-hot encoding methods, proving that learned representations effectively capture the nuances of human intent during a crisis.
Conclusion
This study proposes a privacy-preserving alternative for evacuation prediction that does not require sensitive location data. By leveraging pre-disaster web search behavior, local governments can better prepare first response strategies. Future work will focus on generalizability across different disaster types and the prediction of specific evacuation destinations.
Summary updated