A Large-scale Dataset and Benchmark for Commuting Origin-Destination Flow Generation
Authors: Can Rong, Jingtao Ding, Yan Liu, Yong Li
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To bridge this gap, we introduce a large-scale dataset containing commuting OD flows for 3,333 areas including a wide range of urban environments around the United States. Based on that, we benchmark widely used models for commuting OD flow generation. We surprisingly find that the network-based generative models achieve the optimal performance in terms of both precision and generalization ability, which may inspire new research directions of graph generative modeling in this field. |
| Researcher Affiliation | Academia | Can Rong1 Jingtao Ding1 Yan Liu 2 Yong Li1, 1Department of Electronic Engineering, BNRist, Tsinghua University, Beijing, China 2Computer Science Department, University of Southern California, Los Angeles, CA, U.S.A. EMAIL, EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 Training of the Graph Diffusion Model Algorithm 2 OD Matrix Generation through Trained Graph Diffusion Model |
| Open Source Code | Yes | The dataset and benchmark are available at https://github.com/tsinghua-fib-lab/Commuting ODGen-Dataset. |
| Open Datasets | Yes | To address the above issue, we collect data from multiple sources and construct a large-scale dataset containing commuting OD matrices for 3,333 diverse areas around the whole United States (Large Commuing OD)... And our dataset is curated and publicly available, which can be found at https://github.com/ tsinghua-fib-lab/Commuting ODGen-Dataset. |
| Dataset Splits | Yes | All models utilize the ratio of 8:1:1 for dividing the data into training, validation, and test sets. |
| Hardware Specification | Yes | The computational resources we used to conduct the experiments are as follows: a server with a Intel(R) Xeon(R) Platinum 8358P CPU @ 2.60GHz with 128 cores. The server is equipped with 1TB of RAM and 8 NVIDIA A100 GPUs. |
| Software Dependencies | No | The paper mentions several algorithms and optimizers like 'Adam W optimizer' and 'cosine noise scheduler' but does not specify software library versions (e.g., PyTorch 1.x, scikit-learn 1.x). |
| Experiment Setup | Yes | The graph transformer in diffusion models employs 4 layers with each having 32 hidden dimensions. We utilize 250 diffusion steps in diffusion models, following a cosine noise scheduler as suggested by Nichol & Dhariwal (2021). Denoising networks are optimized using Adam W optimizer (Loshchilov & Hutter, 2017), with a learning rate set at 1e-3. Our method and Diff ODGen both sample 50 times during generation and take the average as final generated results. For the gravity model, we adopt the approach outlined by Barbosa et al. (2018), which involves four fitting parameters. In the random forest algorithm, the number of estimators is set to 100. The DGM (Simini et al., 2021) is stacked by 10 layers with 64 hidden dimensions in each layer, while GNN-based models are designed with 3 layers and 64 channels all. Trans Flower is stacked by 3 transformer encoder with 8 heads and 64 hidden dimensions in each head. The hyper-parameters for the denoising networks in two cascaded diffusion models of Diff ODGen are aligned with our methodology. |