IJCAI25(accepted)

This is the official implementation of our IJCAI25 submission: Unveiling the Power of Noise Priors: Enhancing Diffusion Models for Mobile Traffic Prediction.

---

The repo currently includes code implementations for the following tasks:

12-12 Prediction: We provide the script for the reproduction of 12-12 prediction results in this repo.

12-1 Prediction: We provide the script for the reproduction of 12-1 prediction results in this repo.

Introduction

In this paper, we introduce a novel perspective by emphasizing the role of noise in the denoising process. By incorporating the intrinsic dynamics of mobile traffic data as noise priors, we build NPDiff, a general noise prior framework for diffusion model-based mobile traffic prediction.

Overall Architecture

NPDiff consists of two main parts: (i) extraction of two key dynamics inherent in mobile traffic data, (ii) the denoising process with noise priors.

Data

We use four mobile traffic datasets to demonstrate the effectiveness of NPDiff. You can find the raw datasets in the ./data/mobile_npy/ directory. Before running the experiments, you need to pre-process all data via:

cd /path/to/your/project
bash scripts/data_prepare.sh

⚙️ Installation

Environment

• Tested OS: Linux
• Python >= 3.9
• torch == 2.3.1

Dependencies:

1. Install Pytorch with the correct CUDA version.
2. Use the pip install -r requirements.txt command to install all of the Python modules and packages used in this project.

Model Training

We provide the scripts under the folder ./scripts/. You can train NPDiff with the MobileBJ dataset as the following examples:

python main.py --batch_size 8 --device "cuda:2" --Num_Comp 0 --Lambda 0.5 --model "CSDI" --target_dim 672 --history_len 12 --predict_len 12 --data_name "MobileBJ" 

Once your model is trained, you will find the trained model in the ./save/ directory.

There are some new parameters to specify:

• history_len specifies the input sequence length.
• predict_len specifies the prediction horizon.
• model specifies different denoising networks.
• Num_Comp specifies the number of periodic dynamic components.
• Lambda specifies noise fusion coefficient.