Predicting Human Mobility during Extreme Events via LLM-Enhanced Cross-City Learning

1. Problem Definition

The paper addresses the critical challenge of predicting human mobility during extreme events (e.g., severe weather, natural disasters).

• Context: Rapid urbanization and climate change have increased city vulnerability. Accurate mobility prediction is essential for downstream tasks like early disaster warnings and resource pre-allocation.
• The Gap: Existing mobility prediction models (e.g., DeepMove, Flashback) are trained on normal scenarios. They fail significantly in extreme scenarios because human mobility patterns shift drastically (e.g., increased immobility, altered spatial distributions).
• Key Challenges:
  1. Data Scarcity: Specific extreme event data for a target city is often unavailable before the event occurs, making it impossible to train specialized models from scratch.
  2. Domain Shift: Simply transferring knowledge from other cities via standard transfer learning fails because spatial venue distributions differ, and the nature of the shift caused by different extreme events is complex.
  3. Immobility: Extreme events often cause people to stay put (immobility), a behavior poorly captured by models trained on normal movement data.

2. Methodology: X-MLM Framework

The authors propose X-MLM (cross-eXtreme-event Mobility Language Model), a framework that leverages Large Language Models (LLMs) to model human mobility intentions and transfer this knowledge across cities and scenarios. The framework consists of three core modules:

A. RAG-Enhanced Intention Predictor

• Goal: Convert raw trajectory data into an "intention modality" and retrieve relevant external knowledge to handle data sparsity.
• Intention Translator & Predictor: Uses Travel Feature Extraction (POI categories, road networks) to map trajectories into a shared intention space via Transfer Component Analysis (TCA) and clustering.
• Intention-CLIP: A novel contrastive learning module that aligns the intention modality (trajectory features) with the language modality (LLM vocabulary). It uses an attention mechanism based on the LLM's vocabulary table to create intention-related prototypes, enabling the model to output intention features as language embeddings.
• Retrieval (RAG): To address data sparsity in the target city, the system retrieves similar trajectories from a database containing:
  1. Similar intention sequences from other cities during extreme events.
  2. Similar intention sequences from the target city during normal or other extreme scenarios.

B. LLM-Based Intention Refiner

• Goal: Refine the initial intention prediction to account for the specific impact of the extreme event.
• Mechanism: A fine-tuned LLM (using Prefix-Tuning to save parameters) takes the predicted intention and the retrieved reference sequences.
• Hybrid Prompting Strategy:
  1. Intention-Incorporated Prompt: Instead of text-only inputs, the model uses the intention feature sequences directly as embeddings within the prompt.
  2. Chain-of-Thought (CoT): The LLM is guided through a reasoning process: (1) Is the initial prediction correct? (2) If not, is the next state "immobility"? (3) If not, which specific intention is correct?
  3. Event-Level-Aware Soft Prompt: A learnable embedding vector representing the severity of the extreme event (e.g., rainfall level) is prepended to the input to model event heterogeneity.

C. Intention-Modulated Location Predictor

• Goal: Map the refined intention back to a specific geographic location.
• Mechanism: Takes an arbitrary base mobility prediction model (e.g., LSTM, DeepMove) and modulates its output using the refined intention features.
• Modulation Operations: The intention vector is fused with the base model's mobility embedding via element-wise product, concatenation, or attention mechanisms before the final location classification layer.

3. Key Contributions

1. X-MLM Framework: The first framework to utilize LLMs for cross-city transfer learning specifically for extreme event mobility prediction, focusing on the "intention" level rather than raw trajectories.
2. Intention-CLIP: A novel architecture that aligns trajectory features with language semantics, allowing LLMs to understand and predict mobility intentions.
3. Hybrid Prompting & CoT: A specialized prompting strategy that combines direct intention embeddings with Chain-of-Thought reasoning to explicitly handle the shift toward immobility during disasters.
4. Cross-City Knowledge Transfer: Demonstrates a method to effectively transfer knowledge of how extreme events alter human behavior from source cities to a target city with limited data.

4. Experimental Results

The model was evaluated on 7 real-world datasets from Chinese cities involving heavy rainfall events.

• Performance Gains: Compared to state-of-the-art baselines (including DeepMove, Flashback, CHAML, and LLM-based models like LLM4POI), X-MLM achieved:
  • 32.8% improvement in Acc@1.
  • 28.3% improvement in Mean Reciprocal Rank (MRR).
  • 35.0% improvement in F1-score for predicting immobility (a critical metric for disaster response).
• Robustness: When applied to various base models (LSTM, GRU, DeepMove), X-MLM consistently compensated for the performance drop seen in extreme scenarios, often restoring or exceeding normal-scenario performance levels.
• Ablation Studies: Removing any component (RAG, Soft Prompt, Immobile modeling, or LLM Refining) resulted in significant performance degradation, confirming the necessity of each module.
• Efficiency: While slightly slower than non-LLM baselines, the inference time (approx. 5.6s) is comparable to other LLM-based mobility models, proving practical feasibility.

5. Significance

• Disaster Response: The ability to accurately predict immobility and location shifts during extreme events allows for more effective early warning systems and optimized allocation of rescue resources.
• Methodological Advancement: The paper bridges the gap between Deep Learning for mobility and Large Language Models, demonstrating that LLMs can effectively model complex, non-linear shifts in human behavior caused by external shocks.
• Data Efficiency: By leveraging cross-city knowledge and intention-level abstraction, the framework solves the "cold start" problem for cities facing unprecedented extreme events where historical data is scarce.