Predicting Human Mobility in Disasters via LLM-Enhanced Cross-City Learning
Yinzhou Tang, Huandong Wang, Xiaochen Fan, Yong Li · 2025
Read Predicting Human Mobility in Disasters via LLM-Enhanced Cross-City Learning by Yinzhou Tang, Huandong Wang, Xiaochen Fan, Yong Li (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-19737.txt.
Introduction
Yinzhou Tang, Huandong Wang, Xiaochen Fan, and Yong Li present a framework for forecasting how people move during extreme events, in work from 2025. Floods, storms, and heat waves shift travel in ways normal models miss. Good forecasts feed early warnings and rescue planning.
Their system adds a language model layer on top of existing movement predictors. One part forecasts a person's next intent from past moves and event facts, a second refines that intent, and a third turns intent into a place. Knowledge of how events bend intent is shared across cities, so a model trained in one city can help in another. Tests showed clear gains over base predictors on top-hit and ranking scores.
Catalog record
| CREATOR | Yinzhou Tang, Huandong Wang, Xiaochen Fan, Yong Li |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 53 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-19737
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Predicting Human Mobility in Disasters via LLM-Enhanced Cross-City Learning Yinzhou Tang, Huandong Wang, Xiaochen Fan, Yong Li Abstract The vulnerability of cities to natural disasters has increased with urbanization and climate change, making it more important to predict human mobility in the disaster scenarios for downstream tasks including location-based early disaster warning and pre-allocating rescue resources, etc. However, existing human mobility prediction models are mainly designed for normal scenarios, and fail to adapt to disaster scenarios due to the shift of human mobility patterns under disaster. To address this issue, we introduce DisasterMobLLM , a mobility prediction framework for disaster scenarios that can be integrated into existing deep mobility prediction methods by leveraging LLMs to model the mobility intention and transferring the common knowledge of how different disasters affect mobility intentions between cities. This framework utilizes a RAG-Enhanced Intention Predictor to forecast the next intention, refines it with an LLM-based Intention Refiner, and then maps the intention to an exact location using an Intention-Modulated Location Predictor. Extensive experiments illustrate that DisasterMobLLM can achieve a 32.8% improvement in terms of Acc@1 and a 35.0% improvement in terms of the F1-score of predicting immobility compared to the baselines. The code is available at https://github.com/tsinghua-fib-lab/DisasterMobLLM . I Introduction With the rapid urbanization [ 1 ] and climate change [ 2 ] , cities across the world are becoming increasingly vulnerable to natural disasters (e.g., heavy rains), exposing more human lives and properties at risk. To tackle these challenges, a fundamental research problem is to predict human mobility during disaster scenarios, which can support a wide spectrum of downstream emergency response tasks including location-based early disaster warning [ 3 , 4 , 5 ] , pre-allocating rescue resources [ 6 ] , and planning humanitarian relief [ 7 ] , etc. As a classic machine learning problem, human mobility prediction has been studied for decades; however, most existing work [ 8 , 9 ] has focused on normal scenarios rather than disaster scenarios. As illustrated in Fig. 1 (a) and (b), we employ two representative algorithms trained in the normal scenario, i.e., DeepMove [ 8 ] and Flashback [ 10 ] , to predict human mobility in normal scenarios and disaster scenarios, respectively. Their performance in disaster scenarios significantly decreases compared with normal scenarios, with an average relative performance gap of 46.4% and 24.5% in terms of accuracy and mean reciprocal rank, respectively. This underperformance arises from the significant difference in mobility patterns between normal and disaster scenarios in terms of spatial distribution (see Fig. 1 (c)) and immobility probability [ 11 ] (see Fig. 1 (d)). Therefore, mobility prediction in disaster scenarios remains an open problem. a Accuracy of mobility prediction in normal/disaster scenarios b Mean reciprocal rank of mobility predictions c JSD between daily location distribution and average location distribution vs. precipitation d The number of immobility records in each day vs. precipitation Fig. 1 : Analysis of mobility prediction performance of existing algorithms trained in the normal scenario (i.e., DeepMove and Flashback) as well as cross-city transfer learning algorithm (i.e., CHAML), and the cause of their underperformance. However, developing a specialized mobility prediction model for disaster scenarios is also a challenging task. The most critical issue lies in the diversity of disaster situations, e.g., different rainfalls ranging from once-in-a-decade to once-in-a-century, which are usually beyond our experience. It indicates the lack of available mobility data for a specific disaster in a particular city to train the specialized model before the disaster [ 12 ] . A promising solution to this problem is to leverage data from similar disasters in other cities, transferring knowledge to train the disaster-specialized model effectively. However, we also observed that simply utilizing mobility data of similar disasters from other cities through transfer learning techniques still cannot achieve satisfactory performance. Specifically, the CHAML algorithm [ 13 ] still exhibits a substantial relative performance gap of 38.9% and 31.4% under disaster scenarios compared to normal scenarios in terms of accuracy and mean reciprocal rank (see Fig. 1 (a)). The largely different distribution of spatial venues in source cities and target cities makes it difficult to effectively transfer knowledge. Further, existing cross-city mobility prediction algorithms [ 14 , 15 ] also fail to utilize the massive trajectory data in normal scenarios of source cities. An ideal approach should effectively utilize trajectory data from the source in both normal and disaster scenarios while decoupling distribution differences of spatial venues to enable efficient transfer of common knowledge describing the shift of mobility patterns (e.g., from the perspective of mobility intention) between normal scenarios and disaster scenarios. In this paper, we propose DisasterMobLLM , an LLM-based cross-city transfer learning framework for mobility prediction in disaster scenarios. In this framework, we leverage LLMs to transfer knowledge about how disasters influence human mobility at the intention level, facilitating the understanding and prediction of mobility intentions across different cities in disaster scenarios. Specifically, we propose an RAG-Enhanced Intention Predictor to retrieve similar trajectories with the target user and target disaster situations as external knowledge. We also propose an intention translator and predictor that aligns intention modality and language modality to enhance LLMs to better forecast the subsequent intention. In the subsequently introduced LLM-based Intention Refiner, we utilize an intention-incorporated prompt based on Chain-of-Thought (CoT) reasoning and a disaster-level-awaring soft prompt to refine the predicted intentions incorporating the shift of intention from normal scenarios to disaster scenarios. Finally, we introduce an Intention-Modulated Location Predictor combining an arbitrary base mobility prediction model and intention information to predict the next location of user movement in disaster scenarios. Our contributions can be summarized as follows: • We propose DisasterMobLLM, an LLM-based cross-city transfer learning framework, which utilizes LLM-adapted intention embeddings to modulate arbitrary base models to predict human mobility in disaster scenarios. It uses LLM to understand how different disasters affect human mobility and transfer it across cities and scenarios. • We design an Intention-CLIP to achieve the modality alignment between intention modality and language modality to enhance the ability to understand human mobility intention for LLMs and predict the next intention. • Extensive experiments demonstrate that, compared to baselines, DisatserMobLLM achieve a 32.8% improvement in Acc@1, a 28.3% improvement in MRR, and a 35.0% improvement in the F1-score for immobility, effectively compensating for the performance decrease in disaster scenarios of base models in normal scenarios. II Related Works Traditional Deep Learning Models in Mobility Prediction Since mobility prediction tasks have been widely studied in recent years [ 16 , 9 , 17 ] , there are many mobility prediction models using deep learning methods. Feng et. al. proposed DeepMove [ 8 ] to use RNN and attention to extract time and location embeddings for trajectory prediction. Flashback [ 10 ] enhances the prediction performance of RNNs by emphasizing periodic and contextual information. STiSAN+ [ 18 ] further refines predictions by considering both daily and hourly timing along with spatial relationships. SSDL [ 19 ] represents the complex human mobility semantics in different hidden spaces by decoupling the time-invariant and time-varying factors, and the POI graph structure is designed to explore the heterogeneous cooperation signals in the historical mobility. STAR [ 20 ] captures spatio-temporal correspondence by designing different spatio-temporal