DeepGraphGenerator

Awesome Deep Graph Generator

Source codes implementation of papers:

• BTGAE: Divide and Conquer: A Topological Heterogeneity-based Framework for Scalable and Realistic Graph Generation. (Official PyTorch Implementation)
• VRDAG: Efficient Dynamic Attributed Graph Generation, in ICDE 2025.
• TGAE: Efficient Learning-based Graph Simulation for Temporal Graphs, in ICDE 2025.
• CPGAE: Efficient Learning-based Community-Preserving Graph Generation, in ICDE 2022. (GAE version of CPGAN)

Implementation of baselines:

• GraphRNN: GraphRNN: Generating Realistic Graphs with Deep Auto-regressive Models, in ICML 2018.

Usage

Data processing

1. Run python experiment/preprocess.py to process a graph into a sparse matrix and save the matrix as an .npz file or .pkl file.

Training

To test implementations of the methods, run

python main.py --method BTGAE
python main.py --method VRDAG
python main.py --method TGAE
python main.py --method CPGAE
python main.py --method GraphRNN
python main.py --method GraphRNN-S

All predefined configurations are in config/. Dynamically override parameters using:

python train.py --method <name> --update <key1>=<value1> <key2>=<value2> ...

Key values shall exactly match (case-sensitive) the corresponding parameters defined in configuration files.

Examples:

python main.py --method BTGAE --update data=cora epochs=150 learning_rate=3e-3

python main.py --method TGAE --update epochs=100 lr=1e-3

Evaluaion

The evaluation tools are located in the experiment/graph_metrics directory. For usage instructions, see README.md.

Data Description

The following datasets are mainly from linqs and snap.

$d_{mean}$: mean degree.

GINI: GINI index, which is a common measure for inequality in a degree distribution.

PWE: power-law exponent.

The following temporal network datasets are from snap and Network Repository. For more information, please refer to the VRDAG paper.

Test Result

The performance of models tested on static graph datasets are listed as follows:

The second table shows the test results of dynamic graph generation methods.

The evaluation metrics used in these tests are as follows:

• Deg. dist.: It measures the degree distribution similarity between the generated graph and the real graph using Maximum Mean Discrepancy (MMD). A smaller MMD value indicates a closer degree distribution.

• Clus. dist.: This metric focuses on the clustering coefficient distribution similarity using MMD.

• Wedge count: It counts the number of wedges in the graph.

• Triangle count: The number of triangles in the graph is counted.

• LCC: It represents the size of the largest connected component in the graph.

• PLE: It is the power-law exponent associated with the degree distribution of the graph.

• Gini: GINI index, which is a common measure for inequality in a degree distribution.

• Clus. coef.: The global clustering coefficient of the graph.

For temporal graphs, given a metric $f_m(\cdot)$, the real graph $\widetilde{G}$, and the synthetic one $\widetilde{G^{\prime}}$, we construct a sequence of snapshots $\widetilde{S}^t$ ($\widetilde{S^{\prime}}^t$), $t = 1, ...,T$, of $\widetilde{G}$ ($\widetilde{G^{\prime}}$) by aggregating edges from the initial timestamp to the current timestamp $t$. Then, we measure the average difference (in percentage) of the given metric $f_m(\cdot)$ between two graphs as follows:

$$ f_{avg}(\widetilde{G},\widetilde{G^{\prime}},f_m)=Mean_{t=1:T}(|\frac{f_m(\widetilde{S}^t)-f_m(\widetilde{S^{\prime}}^t)}{f_m(\widetilde{S}^t)}|) $$

Repo Structure

The repository is organized as follows:

• main.py: organize all models.
• experiment/: preprocessing and evaluation.
• methods/: implementations of models.
• config/: configuration files for different models.
• models/: the checkpoints or the trained models for each method.
• data/: dataset files.
• requirements.txt: package dependencies.

Requirements

torch           2.3.1+cu121
networkx        2.8
scipy           1.14.1
scikit-learn    1.5.2
numpy           1.26.4
community       1.0.0b1
python-louvain  0.16
dgl             2.4.0+cu121

Contributors :

Citing

If you find GraphGenerator is useful for your research, please consider citing the following papers:

@inproceedings{li2025efficient,
  title={Efficient dynamic attributed graph generation},
  author={Li, Fan and Wang, Xiaoyang and Cheng, Dawei and Chen, Cong and Zhang, Ying and Lin, Xuemin},
  booktitle={2025 IEEE 41st International Conference on Data Engineering (ICDE)},
  pages={1415--1428},
  year={2025},
  organization={IEEE}
}

@inproceedings{xiang2025efficient,
  title={Efficient learning-based graph simulation for temporal graphs},
  author={Xiang, Sheng and Xu, Chenhao and Cheng, Dawei and Wang, Xiaoyang and Zhang, Ying},
  booktitle={2025 IEEE 41st International Conference on Data Engineering (ICDE)},
  pages={251--264},
  year={2025},
  organization={IEEE}
}

@inproceedings{xiang2022efficient,
  title={Efficient learning-based community-preserving graph generation},
  author={Xiang, Sheng and Cheng, Dawei and Zhang, Jianfu and Ma, Zhenwei and Wang, Xiaoyang and Zhang, Ying},
  booktitle={2022 IEEE 38th International Conference on Data Engineering (ICDE)},
  pages={1982--1994},
  year={2022},
  organization={IEEE}
}

@article{xiang2022general,
  title={General graph generators: experiments, analyses, and improvements},
  author={Xiang, Sheng and Wen, Dong and Cheng, Dawei and Zhang, Ying and Qin, Lu and Qian, Zhengping and Lin, Xuemin},
  journal={The VLDB Journal},
  pages={1--29},
  year={2022},
  publisher={Springer}
}