📘 Workflow of Data Generation

This README describes the complete workflow of urban mobile traffic and mobile user distribution generation, including dataset preprocessing, model training, and model inference. Please execute all commands in the following directory:

./ZoomDiff/

Or execute directly in the current directory:

cd ./ZoomDiff/

---

🗂️ Dataset Preprocessing

1️⃣ Data Preparation

1.1 Population Distribution Data

Please place the national population distribution data under the following directory:

./datasets/data_preparation/geographic_data/

Taking China as an example, the required geographic data includes:

• 🗺️ China Population Distribution file (.tif) Download from: https://hub.worldpop.org/geodata/listing?id=135

1.2 Geographic Data of the Target City

Please place the geographic data of the target city under the following directory:

./datasets/data_preparation/geographic_data/

Taking Chengdu as an example, the required geographic data includes:

• 🗺️ OSM – Sichuan Province Download from: https://download.geofabrik.de/asia/china.html
• 📍 POI – Chengdu You may obtain POI data from any suitable source. ⚠️ Important: The POI data used for training and inference must share the same set of bigtype categories and identical encodings.

---

2️⃣ Preprocess City Data for Training

To preprocess city data for model training, execute the following scripts in order:

• 📦 Grid-level Data Transformation

  python ./datasets/data_preparation/Grid_meta.py \
    --cityname "${CITYNAME}" \
    --province "${PROVINCE}" \
    --data_path "${DATA_PATH}" \
    --shp_path "${SHP_PATH}" \
    --pop_path "${POP_PATH}"
  

  Please fill in the correct names or paths in ${...}. Alternatively, you can directly run:

  ./datasets/data_preparation/src/Grid_meta_gen.sh
  

• ⚙️ Environmental Feature Fusion

  python ./datasets/data_preparation/Env_extra.py \
    --cityname "${CITYNAME}" \
    --province "${PROVINCE}" \
    --shp_path "${SHP_PATH}" \
    --pop_path "${POP_PATH}" \
    --osm_dir "${OSM_DIR}" \
    --poi_dir "${POI_DIR}"
  

  Please fill in the correct names or paths in ${...}. Alternatively, you can directly run:

  ./datasets/data_preparation/src/Env_gen.sh
  

---

3️⃣ Preprocess City Data for Inference

For model inference, only the condition preparation step is required:

• ⚙️ Environmental Feature Fusion

  ./datasets/data_preparation/src/Env_gen.sh
  

---

🚀 ZoomDiff Model Training

For the detail of ZoomDiff model, please refer to Denoising Refinement Diffusion Models for Simultaneous Generation of Multi-scale Mobile Traffic：

X. Qi, H. Chai, S. Liu, L. Yue, R. Pan, Y. Wang, and Y. Li, “Denoising refinement diffusion models for simultaneous generation of multi-scale mobile network traffic,” arXiv preprint arXiv:2511.17532, Oct. 2025, doi: 10.48550/arXiv.2511.17532.

Run the following command to train the ZoomDiff model:

python ZoomDiff_train.py \
  --dataset TrainCity1*TrainCity2*TrainCity3*... \
  --datatype traffic \
  --device cuda:0

Arguments:

• 🏙️ --dataset: Training cities, separated by *
• 📊 --datatype: Type of data (e.g., traffic, user)
• 💻 --device: Computing device (e.g., cpu, cuda:0)

---

🔍 ZoomDiff Model Inference

Run the following command to perform inference:

python ZoomDiff_infer.py \
  --dataset InfCity1*InfCity2*InfCity3*... \
  --datatype traffic \
  --device cuda:0 \
  --nsample 1

Arguments:

• 🏙️ --dataset: Target cities for inference, separated by *
• 📊 --datatype: Type of data (e.g., traffic, user)
• 💻 --device: Computing device (e.g., cpu, cuda:0)
• 🎯 --nsample: Number of samples generated per single inference run

---

✨ This pipeline ensures consistent preprocessing, reliable training, and reproducible inference across multiple cities.