🤖 AI

EventOD: Event-Aware OD Flow Generation via LLM-Guided Semantic Modulation

EventOD is an event-adaptive framework that leverages large language models to derive semantic control vectors and lightweight calibration modules to steer a pretrained graph diffusion OD generator, enabling accurate origin-destination flow estimation under disruptive events without retraining.

Original authors: Jie Zhao, Jie Feng, Can Rong, Zhihan Hou, Peng Lu, Yong Li

Published 2026-07-28

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Jie Zhao, Jie Feng, Can Rong, Zhihan Hou, Peng Lu, Yong Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to predict how a city breathes. In the quiet, boring days of a normal week, people move in predictable patterns: they wake up, commute to work, grab lunch, and head home. Scientists call this "Origin-Destination" (OD) flow, which is just a fancy map showing where people start their trips and where they end up. For years, computers have been getting really good at guessing these routes using deep learning, almost like a GPS that knows your habits better than you do. But here's the catch: what happens when the world suddenly goes crazy? When a hurricane rips through town or a pandemic forces everyone to stay inside, those old habits vanish. The computer's "normal day" map becomes useless because the city's functions have changed—shops are closed, roads are flooded, and people are scared. Trying to teach a computer a brand-new way to predict movement for every single disaster is slow, expensive, and often impossible because we don't have enough data when the crisis hits.

This is where a new idea called EventOD steps in, acting like a smart translator for our city-predicting computers. Instead of throwing away the old, reliable map and trying to learn a new one from scratch, EventOD asks a super-smart AI (a Large Language Model) to read the news about the disaster and figure out how the city's mood has changed. It then gently nudges the old map to fit the new reality without breaking it. Think of it like wearing a pair of "disaster glasses" over your regular glasses; you still see the world clearly, but now you can also see the rain, the wind, and the panic, adjusting your path accordingly. The researchers tested this on real hurricanes and pandemics across the United States, and they found that this "gentle nudge" approach is much better at predicting where people will go during a crisis than trying to force the computer to relearn everything from scratch.

The Problem: When the City Breaks Its Routine

You know how your phone's weather app is great at predicting rain on a Tuesday, but it might get confused if a tornado suddenly touches down? That's exactly what happens with computer models that predict human movement. These models are trained on "routine" data—how people move when everything is normal. They learn that people go to schools in the morning and parks in the afternoon. But when a disruptive event hits, like a massive storm or a global health crisis, the rules change instantly. A school might become a shelter, or a park might become a no-go zone.

The problem is that these computers are stubborn. If you try to teach them the new rules during a disaster, you usually don't have enough data to do it properly. The event is happening now, and the data is messy or missing. If you try to retrain the whole computer model for every single hurricane or pandemic, it's like trying to rebuild a car engine while it's still driving down the highway. It's too slow, and you might break the engine. The researchers realized that instead of rebuilding the engine, they just needed to adjust the steering wheel.

The Solution: The "Semantic" Steering Wheel

The team behind EventOD came up with a clever three-step plan to fix this without breaking the computer's brain.

Step 1: The Translator (The LLM)
First, they use a Large Language Model (LLM)—the same kind of AI that writes essays or chats with you—to read the "news" of the disaster. They feed the AI information about the event, like wind speeds for a hurricane or infection rates for a pandemic. The AI doesn't just say "it's bad"; it translates that chaos into a structured list of changes. It figures out, for example, that "restaurants will likely be closed" (a functional change) and "families with kids will stay home" (a demographic change). It turns this into a simple control vector, a set of instructions like "decrease activity here, increase activity there."

Step 2: The Volume Knobs (AlphaNet and BetaNet)
Knowing what to change is great, but knowing how much to change is even better. The AI's instructions are just directions (like "turn left"), but the computer needs to know the speed of the turn. This is where two tiny, lightweight computer programs called AlphaNet and BetaNet come in. They act like volume knobs. AlphaNet adjusts the "functional" volume (how much the shops and offices are affected), and BetaNet adjusts the "demographic" volume (how much different groups of people are affected). These knobs are learned from a little bit of data, allowing the system to fine-tune the AI's rough guesses into precise adjustments.

Step 3: The Safety Net (Retrieval-Augmented Fallback)
Sometimes, the data is so sparse that even the volume knobs get confused. To handle this, EventOD has a backup plan. It looks back at its training library to find similar past disasters. If it's a new type of storm, it asks, "Hey, how did we handle a similar storm before?" It uses this memory to fill in the gaps, ensuring the system doesn't crash when the data is thin.

How It Works in Practice

Once the system has its "control vectors" (the directions) and its "knobs" (the magnitude), it injects this information directly into the frozen, pre-trained computer model. The original model stays exactly the same—its brain isn't changed at all. Instead, the input data is slightly tweaked to reflect the disaster. It's like giving a seasoned driver a new set of instructions: "The road is flooded, so take the detour," without needing to retrain the driver on how to drive.

The researchers tested this on two very different scenarios: Hurricane Dorian in 2019, which caused massive physical destruction and weather-driven chaos, and the COVID-19 pandemic in 2020, which caused behavioral and policy-driven changes. In both cases, they compared EventOD against other methods, including trying to retrain the model from scratch or just using the old model without any changes.

The Results: Smarter, Faster, and More Accurate

The results were clear. EventOD consistently outperformed the other methods. In the hurricane scenario, it improved the accuracy of predicting where people would go by a significant margin. Specifically, it reduced the error in predicting the flow of people by about 18% compared to the best pre-trained model that hadn't been adapted. It also did a much better job at matching the actual distribution of people, meaning it didn't just guess the right number of people, but guessed the right places they would go.

One of the most interesting findings was that keeping the original model frozen was actually better than trying to fine-tune it. When the researchers tried to retrain the model on the limited disaster data, it actually performed worse. This suggests that the "common sense" the model learned from normal days is valuable, and you don't want to overwrite it with noisy, incomplete disaster data. Instead, you just need to gently steer it.

The system also proved to be flexible. It worked well on different types of computer models (not just the one they started with) and handled different kinds of disasters, from weather shocks to policy restrictions. Even when they tested it on a "hard mode" where they gave it very little training data, it didn't crash; it just got a little less accurate, but still far better than the alternatives.

The Cost of Being Smart

You might wonder, "Does using a super-smart AI to translate the disaster news cost a fortune or take forever?" The researchers checked this too. They found that the process is surprisingly cheap and fast. For each region (like a county), the cost to run the AI translation and adjustment was less than $0.002. The time it took was dominated by the AI thinking, taking a few seconds per region, but since these calculations can be done in parallel for many regions at once, the total time is manageable. The "retrieval" part, where the system looks up past disasters, was almost instant, taking only about 0.0001 seconds.

Why This Matters

EventOD shows us that when the world goes off the rails, we don't need to throw away our old tools. We just need to teach them how to read the new situation. By combining the deep knowledge of "normal life" that computers already have with the flexible, language-understanding power of modern AI, we can create systems that are ready for the unexpected. Whether it's a hurricane knocking out power lines or a pandemic changing how we socialize, this approach offers a way to keep our cities moving—and our emergency responders informed—without starting from zero every time the sky turns gray.

Technical Summary of EventOD: Event-Aware OD Flow Generation via LLM-Guided Semantic Modulation

Problem Statement
Estimating Origin-Destination (OD) flows during disruptive events (e.g., hurricanes, pandemics) is critical for disaster response and urban resilience. However, existing deep OD models, typically trained on routine mobility data, suffer significant performance degradation when extreme events abruptly alter regional functions and population activities. These models face a "distribution shift" where historical priors no longer align with event-time conditions. Retraining a new generator for every unique event is impractical due to the scarcity of event-time supervision and the heterogeneity of disruptive scenarios. Furthermore, direct fine-tuning of pretrained generators on sparse event data can distort the stable spatial interaction patterns learned during pretraining. The core challenge is to adapt a frozen, pretrained OD generator to disruptive contexts using only coarse event signals (e.g., weather data, policy descriptions) and limited supervision, while preserving the model's structural fidelity.

Methodology: The EventOD Framework
EventOD addresses this challenge by treating event adaptation as a semantic modulation problem rather than a parameter relearning task. The framework operates in four stages:

LLM-based Impact Semantic Extraction:
Instead of using raw event data directly, the framework employs a Large Language Model (LLM) to infer structured semantic control vectors from coarse event observations.
- POI Control Vector ( $v^{(p)}_i$ ): The LLM analyzes regional Point of Interest (POI) distributions and event context to generate a discrete vector (values $\{-1, 0, 1\}$ ) indicating the expected directional change (decrease, unchanged, increase) in functional activity for each POI category.
- Demographic Control Vector ( $v^{(d)}_i$ ): The LLM infers behavioral adjustments across demographic groups based on the demographic attributes and the inferred functional shifts.
  These vectors serve as high-level semantic priors characterizing how the event impacts urban supply (functions) and demand (demographics).
Data-driven Impact Factor Learning:
Since LLMs provide direction but not magnitude, EventOD introduces two lightweight neural networks, AlphaNet and BetaNet, to learn continuous control factors ( $\alpha_i, \beta_i$ ) that quantify the intensity of these semantic shifts.
- AlphaNet learns functional control factors by combining regional POI features with auxiliary event descriptors (e.g., meteorological variables).
- BetaNet learns demographic control factors conditioned on demographic attributes and the POI control vector, ensuring demographic adaptation is driven by event-conditioned functional changes rather than raw exogenous noise.
  These factors modulate the original features in log-space: $\tilde{p}_i = p_i \odot \exp(\alpha_i \odot v^{(p)}_i)$ and $\tilde{d}_i = d_i \odot \exp(\beta_i \odot v^{(d)}_i)$ .
Knowledge-guided Fallback (Retrieval-Augmented Reasoning):
To handle scenarios with sparse supervision or weakly represented event types, EventOD includes a retrieval-augmented pathway. It constructs knowledge banks (AlphaKnow, BetaKnow) from training data. When data-driven networks are unreliable, an LLM retrieves similar historical control factor distributions and generates complementary estimates via semantic reasoning, serving as a robust fallback without test-time information leakage.
Event-Conditioned Graph Diffusion:
The modulated regional features ( $\tilde{x}_i$ ) are injected into a frozen pretrained graph diffusion OD generator (ODNet). The generator's parameters remain fixed; only the input conditioning features are altered. This allows the model to adapt to event-specific dynamics without retraining the core generator. The authors provide a theoretical analysis showing that this input-level modulation introduces bounded local perturbations to the reverse diffusion process, ensuring stability.

Key Contributions

Problem Formulation: The paper formulates event-adaptive OD generation as a practical mobility adaptation problem, aiming to recover event-conditioned flows from coarse context while preserving pretrained priors.
Framework Design: EventOD proposes a novel semantic adaptation framework that combines LLM-derived control vectors, lightweight factor learning (AlphaNet/BetaNet), and frozen-generator modulation. This enables event-aware generation without event-specific retraining.
Robustness Mechanism: The introduction of a retrieval-augmented fallback pathway enhances robustness in data-sparse scenarios.
Empirical Validation: Extensive experiments across hurricane and pandemic scenarios demonstrate that EventOD improves both reconstruction accuracy and distributional fidelity compared to strong baselines.

Experimental Results
The framework was evaluated on two distinct scenarios:

Hurricane Dorian (2019): Using 184 U.S. counties with meteorological data.
COVID-19 Pandemic (2020): Using 200 U.S. counties with policy and epidemic signals.

Key Findings:

Performance: EventOD consistently outperformed classical models (Gravity, Radiation), conventional ML (RF, SVR), and deep learning baselines (DGM, GMEL). It also surpassed a fine-tuned version of the pretrained generator (WeDANft), achieving a 32.3% improvement in Common Part of Commuting (CPC) and an 18.0% reduction in RMSE over the frozen baseline in the hurricane scenario.
Ablation: Removing event-aware modulation resulted in significant performance drops. Both functional (Alpha) and demographic (Beta) modulation were necessary for optimal results, with functional modulation driving flow overlap and demographic modulation preserving distributional structure.
Generalization: The framework transferred effectively to the pandemic scenario, where demographic modulation proved relatively more important than in the hurricane scenario.
Portability: EventOD successfully improved performance on both diffusion-based (WeDAN) and GAN-based (NetGAN) pretrained generators, demonstrating its independence from specific generator architectures.
Control Quality: Experiments confirmed that LLM-derived semantic controls significantly outperformed random or perturbation-based controls, validating the necessity of semantic reasoning.
Efficiency: The computational overhead is dominated by LLM inference (approx. $0.0017 per region), with retrieval being negligible. The process is parallelizable across regions.

Significance and Claims
The paper claims that EventOD offers a practical solution for the "limited supervision" problem in urban intelligence. By decoupling the adaptation mechanism from the generator parameters, it avoids the instability of fine-tuning on sparse data. The authors argue that their approach is superior to parameter-level adaptation because it preserves the structural regularities captured by large-scale pretraining while explicitly modeling the semantic shifts induced by disruptive events. The framework is presented as a modular, interpretable, and robust method for situational awareness and emergency response, capable of adapting to heterogeneous event mechanisms without requiring event-specific retraining.