MoDiff - Graph Generation with Motif-Aware Diffusion Model

Code for the paper MoDiff-Graph Generation with Motif-Aware Diffusion Model (SIGKDD 2025).

Dependencies

MoDiff is built in Python 3.7.12 and Pytorch 1.10.1. Please use the following command to install the requirements:

pip install -r requirements.txt

Running Experiments

Data Preparation

The subgraphs for training are preprocessed with motif detection and stored in different files in this code. We provide StackOverflow as an example. The 1h scale can be used directly, while the 5h and 1k scales need to be decompressed. More data could be found here. We are also working on merging them in one graph, and will update that later.

Train model

CUDA_VISIBLE_DEVICES=${gpu_ids} python main.py --type train_comp --scale ${scale} --config_folder ${folder_name} --config_prefix ${config_prefix} 

for example,

CUDA_VISIBLE_DEVICES=0 python main.py --type train_comp -scale 1h --config_folder Stack --config_prefix SO
python main.py --type train_comp  #Default settings for a quick start

Evaluation model

To generate graphs using the trained score models, run the following command.

CUDA_VISIBLE_DEVICES=${gpu_ids} python main.py --type eval_comp --scale ${scale} --config_folder ${folder_name} --config_prefix ${config_prefix} --ckpt_name ${ckpt_name}

for example,

CUDA_VISIBLE_DEVICES=0 python main.py --type eval_comp -scale 1h --config_folder Stack --config_prefix SO --ckpt_name Sep30_SO1h_comp
python main.py --type eval_comp  #Default settings for a quick start

We provide trained models on StackOverflow for evaluation. Currently, the density of the generated subgraph can be adjusted manually or inferred from the training data. We are also developing a simple module to automatically predict the target density based on historical density patterns.

Citation

If you found the provided code with our paper useful in your work, we kindly request that you cite our work.

ACM Reference Format:

Yuwei Xu and Chenhao Ma. 2025. MoDiff- Graph Generation with Motif-Aware Diffusion Model. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (KDD ’25), August 3–7, 2025, Toronto, ON, Canada. ACM, New York, NY, USA, 12 pages. https://doi.org/10.1145/3711896.3737053

Generation Application

If you're from a non-computer science background and find it challenging to generate usable graph data with this project, feel free to reach out to us through email.

This project (MoDiff) is primarily designed for generating directed graphs like social networks. If you're working with undirected graphs (e.g., molecular structures in biology), you may find this related work from NTU is more suitable: Fast Graph Generation via Spectral Diffusion (IEEE TPAMI 2023). https://github.com/ltz0120/Fast_Graph_Generation_via_Spectral_Diffusion

We’re also always interested in exploring interdisciplinary applications and welcome collaboration opportunities.