Technical Summary: DynaOD

Problem Statement

Dynamic Origin-Destination (OD) flow generation aims to synthesize realistic urban mobility patterns between geographical regions based solely on temporal context and regional attributes, without relying on historical OD observations or real-time traffic measurements. While static OD generation methods exist, they are limited to time-aggregated settings and fail to capture evolving temporal dynamics. Existing dynamic approaches typically depend on historical OD sequences or real-time traffic data, rendering them ineffective in cold-start scenarios or for cross-city transfer where such data is unavailable. The core challenge lies in translating latent semantic temporal signals (e.g., weekdays, holidays) into temporally coherent OD patterns while preserving the inherent spatial heterogeneity of urban regions, all under minimal supervision.

Methodology: DynaOD Framework

DynaOD is a semantic-driven, modular framework designed to condition pretrained static OD generators for dynamic flow synthesis. It operates through three primary stages:

1. Discrete Semantic Modeling of Temporal Context

The framework leverages Large Language Models (LLMs) to infer discrete directional trends from temporal contexts (e.g., "holiday," "weekday"). Instead of predicting numerical flow changes directly, the LLM generates two types of control vectors for each region:

• POI Control Vector ($c_{poi}$): Indicates whether specific Point-of-Interest (POI) categories are expected to increase, decrease, or remain stable.
• Demographic Control Vector ($c_{demo}$): Indicates similar directional shifts for demographic attributes, conditioned on inferred POI trends.
  These vectors take values in $\{-1, 0, 1\}$, representing qualitative shifts in urban activity patterns. This approach separates semantic trend inference from continuous magnitude modeling, leveraging the LLM's commonsense reasoning capabilities.

2. Continuous Temporal Evolution (ShapeNet)

To translate discrete controls into continuous, time-varying region features, DynaOD employs ShapeNet, a differentiable shape generator.

• Input: Static region features (POI and demographic vectors), the discrete control vectors, and time-context encodings.
• Process: ShapeNet outputs Gaussian parameters ($\mu, \sigma^2$) for POI and demographic dynamics. These are sampled to create continuous temporal shapes ($s_{poi}, s_{demo}$).
• Modulation: The static features are modulated multiplicatively using the discrete controls and continuous shapes:
  $$\tilde{x}_{i,t} = x_i \odot (1 + s_{i,t} \odot c_{i,t})$$
  This produces time-varying region features that capture both the direction and magnitude of change over time.

3. Plug-and-Play Generation and Generalization

The modulated features are fed into a frozen, pretrained static OD generator (e.g., WeDAN or NetGAN) to synthesize dynamic OD flows at each time step. This design allows DynaOD to adapt to diverse temporal contexts without retraining the underlying generator.

• Cross-City Generalization (ShapeMem): To handle unseen cities, DynaOD utilizes a retrieval-based memory of composable shape priors. Instead of running ShapeNet inference for new regions, it retrieves and fuses shape priors from similar regions (based on graph embeddings) to approximate temporal dynamics.
• Efficient Inference (Distillation): To reduce the computational cost of querying large LLMs for every region, a lightweight student model (e.g., Qwen-2.5-1.5B) is distilled from a teacher LLM. The student learns to predict the discrete directional control vectors, enabling scalable deployment.

Key Contributions

1. Semantic-Driven Framework: A novel approach that disentangles temporal semantics into discrete directional trends and continuous temporal dynamics, enabling context-aware OD generation under minimal supervision.
2. Retrieval-Enhanced Adaptation: A mechanism (ShapeMem) that reuses regional knowledge across temporal contexts via similarity-based retrieval, significantly improving cross-city generalization and generation fidelity.
3. Modular and Scalable Architecture: A lightweight, plug-and-play design compatible with diverse pretrained OD generators, coupled with an efficient LLM distillation strategy for scalable inference.

Experimental Results

The authors evaluated DynaOD on a large-scale dataset covering 500 U.S. counties with a strict two-axis split protocol (unseen counties and unseen dates).

• Performance: DynaOD consistently outperformed representative baselines, including physics-inspired models (Gravity Model), statistical learning methods (RF, SVR), deep learning models (DGM), and generative graph models (NetGAN, WeDAN).
• Metrics: It achieved a 34.7% improvement in CPC (Common Part of Commuting) over the strongest baseline and reduced distributional divergence (JSD) by up to 42.8%.
• Ablation Studies:
  • Removing directional controls led to significant performance degradation, confirming the necessity of semantic guidance.
  • Retrieval-based shape composition (#Retrieval) outperformed direct ShapeNet inference (#Inference) in unseen settings, demonstrating the efficacy of the memory mechanism.
  • Distilled student models achieved performance close to the teacher LLM while drastically reducing inference costs.
• Modularity: The framework successfully improved performance when applied to different underlying generators (NetGAN and WeDAN), validating its generator-agnostic nature.

Significance and Claims

The paper claims that DynaOD addresses a fundamental gap in urban mobility modeling: the ability to generate dynamic OD flows without historical data or real-time measurements. By unifying semantic temporal control with generative modeling, the framework enables:

• Cold-Start Capability: Generating mobility patterns for new cities or time periods where historical OD data is absent.
• Temporal Coherence: Producing flows that evolve smoothly over time, reflecting realistic human routines and stochastic events.
• Scalability: Offering a deployment strategy that is computationally efficient and adaptable to various pretrained models.

The authors position DynaOD as a step toward more flexible, context-aware urban simulation tools that do not rely on the restrictive assumptions of historical data dependency found in current state-of-the-art methods.