MIDiff

MIDiff: Tackling Sparsity and Imbalance in Mobile Usage Generation via Multivariate-Imaging Diffusion

Official repository for MIDiff, a diffusion-based framework for generating user-level mobile usage traces. It converts sparse multivariate traces into Cross-Gramian Angular Sum Field (C-GASF) images, trains a Triple-Attention U-Net diffusion model in the image space, and converts generated images back to mobile usage sequences.

• Paper: MIDiff: Tackling Sparsity and Imbalance in Mobile Usage Generation via Multivariate-Imaging Diffusion
• Repository: GitHub - YilaiLiu-HKU/MIDiff

<p align="center"> <img src="https://raw.githubusercontent.com/YilaiLiu-HKU/MIDiff/main/pics/framework2_readme.png" width="95%" alt="MIDiff framework"> </p>

Installation

git clone https://github.com/YilaiLiu-HKU/MIDiff.git
cd MIDiff
python -m pip install -r requirements.txt

Download Checkpoint

The paper checkpoint can be downloaded from this Hugging Face repository using the huggingface-cli:

huggingface-cli download YilaiLiu-HKU/MIDiff ckpt/midiff/ema_0.9999_048000.pt --local-dir .

Usage

Please place your processed data under data/ and run the following scripts to sample and evaluate:

1. Sample C-GASF images from the checkpoint:

   bash run_scripts/sample_midiff.sh
   

2. Convert the sampled NPZ into eval-format CSV:

   bash run_scripts/infer_npz_to_eval_csv.sh
   

3. Evaluate generated traces:

   bash run_scripts/evaluate_midiff_real.sh
   

For more details on training and datasets, please refer to the GitHub repository.

Citation

@misc{liu2026midifftacklingsparsityimbalance,
      title={MIDiff: Tackling Sparsity and Imbalance in Mobile Usage Generation via Multivariate-Imaging Diffusion}, 
      author={Yilai Liu and Shiyuan Zhang and Hongyang Du},
      year={2026},
      eprint={2607.14249},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2607.14249}, 
}