A repository for benchmarking continuous-time dynamic graph models for link property prediction.

With the code provided in this repository, we benchmark the performance of several state-of-the-art continuous-time dynamic graph models on transductive link prediction tasks.

This repo utilizes the datasets and evaluation framework of TGB. For further information about TGB, please consult TGB website or its repo.

We benchmark the transductive dynamic link prediction task on the dataset provided by TGB for the dynamic link property prediction. These includes tgbl-wiki, tgbl-review, tgbl-coin, tgbl-comment, and tgbl-flight. A summary of datasets cab be found on TGB Learderboard.

The following continuous-time dynamic graph models can be utilized as TGB baselines for dynamic link property prediction task:

JODIE, DyRep, TGAT, TGN, CAWN, EdgeBank, TCL, GraphMixer, DyGFormer.

For training a model for transductive dynamic link property prediction on a dataset, you can use the following command:

dataset="tgbl-wiki"
model="GraphMixer"

python train_tgb_lpp.py --dataset_name "$dataset" --model_name "$model"

The above command trains and evaluates a GraphMixer model on the tgbl-wiki dataset.

The exact configuration arguments can be found in utils/load_configs.py file.

The required dependencies are specified in the requirements.txt file.

The code is adapted from DyGLib. Thanks to the DyGLib authors for sharing their code. If this code repo is useful for your research, please consider citing the original authors from DyGLib paper as well.

If this repository is helpful for your research, please consider citing our TGB paper below.

@article{huang2023temporal,
  title={Temporal Graph Benchmark for Machine Learning on Temporal Graphs},
  author={Huang, Shenyang and Poursafaei, Farimah and Danovitch, Jacob and Fey, Matthias and Hu, Weihua and Rossi, Emanuele and Leskovec, Jure and Bronstein, Michael and Rabusseau, Guillaume and Rabbany, Reihaneh},
  journal={arXiv preprint arXiv:2307.01026},
  year={2023}
}