Description: This repository contains the reference implementation of CoEvoGNN model proposed in the paper Learning Attribute-Structure Co-Evolutions in Dynamic Graphs accepted by The Second International Workshop on Deep Learning on Graphs: Methods and Applications (DLG-KDD’20) and won the Best Paper Award.
This code package was developed and tested with Python 3.7 and PyTorch 1.0.1.post2.
Make sure all dependencies specified in the ./requirements.txt file are satisfied before running the model. This can be achieved by
pip install -r requirements.txt
Other environment management tool such as Conda can also be used.
The ./data/ folder contains the evolutionary co-authorship graphs (2k and 10k) descripted in the paper. Predefined paths for locating necessary data files can be found in the ./config.py file.
Note: Due to file size constraints, the 10k dataset files are compressed inside ./data/data_10k.zip. Please unzip the file for training on the dataset.
To train the model, run
python main.py --dataset 2k --epochs 10
List of arguments:
--dataset: The dataset of evolutionary co-authorship graphs to use. Valid choices include2kand10k. Default is2k--t_0: The start index of available time points. Default is0--T: Length of available training time points. Default is8--t_train: Length of training time points (from--t_0). Default is8--t_forecast: Number of forecasting snapshots. Default is1--K: Num of layers fusing new time point. Default is2--epochs: Number of epochs for training. Default is10--H_0_npf: File for initializing H_0.
Other examples are provided in the ./example.sh file.
If you find this code pacakage is helpful, please consider cite us:
@article{wang2020learning,
title={Learning Attribute-Structure Co-Evolutions in Dynamic Graphs},
author={Wang, Daheng and Zhang, Zhihan and Ma, Yihong and Zhao, Tong and Jiang, Tianwen and Chawla, Nitesh V and Jiang, Meng},
journal={arXiv preprint arXiv:2007.13004},
year={2020}
}