This repository is built for the paper Towards Better Dynamic Graph Learning: New Architecture and Unified Library.

🔔 If you have any questions or suggestions, please feel free to let us know. You can directly email Le Yu using the email address [email protected] or post an issue on this repository.

• 🔥🔥🔥[2023/09] Our paper is accepted by NeurIPS 2023 (Poster). The camera ready version is coming soon.
• 🔥🔥🔥[2023/09] DyGFormer is evaluated on TGB Leaderboards and it currently ranks first on tgbl-wiki-v2 and tgbl-coin. It also performs good on other datasets. See our other repository DyGLib_TGB for more details.

Dynamic Graph Library (DyGLib) is an open-source toolkit with standard training pipelines, extensible coding interfaces, and comprehensive evaluating strategies, which aims to promote standard, scalable, and reproducible dynamic graph learning research. Diverse benchmark datasets and thorough baselines are involved in DyGLib.

Fourteen datasets are used in DyGLib, including Wikipedia, Reddit, MOOC, LastFM, Myket, Enron, Social Evo., UCI, Flights, Can. Parl., US Legis., UN Trade, UN Vote, and Contact. The first five datasets are bipartite, and the others only contain nodes with a single type.

Most of the used original dynamic graph datasets come from Towards Better Evaluation for Dynamic Link Prediction, which can be downloaded here. Please download them and put them in DG_data folder. The Myket dataset comes from Effect of Choosing Loss Function when Using T-batching for Representation Learning on Dynamic Networks and can be accessed from here. The original and preprocessed files for Myket dataset are included in this repository.

We can run preprocess_data/preprocess_data.py for pre-processing the datasets. For example, to preprocess the Wikipedia dataset, we can run the following commands:

cd preprocess_data/
python preprocess_data.py  --dataset_name wikipedia

We can also run the following commands to preprocess all the original datasets at once:

cd preprocess_data/
python preprocess_all_data.py

Eight popular continuous-time dynamic graph learning methods are included in DyGLib, including JODIE, DyRep, TGAT, TGN, CAWN, EdgeBank, TCL, and GraphMixer. Our recent work DyGFormer is also integrated into DyGLib, which can explore the correlations of the source node and destination node by a neighbor co-occurrence encoding scheme, and effectively and efficiently benefit from longer histories via a patching technique.

DyGLib supports dynamic link prediction under both transductive and inductive settings with three (i.e., random, historical, and inductive) negative sampling strategies, as well as dynamic node classification.

New datasets and new models are welcomed to be incorporated into DyGLib by pull requests.

• For new datasets: The format of new datasets should satisfy the requirements in DG_data/DATASETS_README.md. Users can put the new datasets in DG_data folder, and then run preprocess_data/preprocess_data.py to get the processed datasets.
• For new models: Users can put the model implementation in models folder, and then create the model in train_xxx.py or evaluate_xxx.py to run the model.

PyTorch 1.8.1, numpy, pandas, tqdm, and tabulate

Dynamic link prediction could be performed on all the thirteen datasets. If you want to load the best model configurations determined by the grid search, please set the load_best_configs argument to True.

• Example of training DyGFormer on Wikipedia dataset:

python train_link_prediction.py --dataset_name wikipedia --model_name DyGFormer --patch_size 2 --max_input_sequence_length 64 --num_runs 5 --gpu 0

• If you want to use the best model configurations to train DyGFormer on Wikipedia dataset, run

python train_link_prediction.py --dataset_name wikipedia --model_name DyGFormer --load_best_configs --num_runs 5 --gpu 0

Three (i.e., random, historical, and inductive) negative sampling strategies can be used for model evaluation.

• Example of evaluating DyGFormer with random negative sampling strategy on Wikipedia dataset:

python evaluate_link_prediction.py --dataset_name wikipedia --model_name DyGFormer --patch_size 2 --max_input_sequence_length 64 --negative_sample_strategy random --num_runs 5 --gpu 0

• If you want to use the best model configurations to evaluate DyGFormer with random negative sampling strategy on Wikipedia dataset, run

python evaluate_link_prediction.py --dataset_name wikipedia --model_name DyGFormer --negative_sample_strategy random --load_best_configs --num_runs 5 --gpu 0

Dynamic node classification could be performed on Wikipedia and Reddit (the only two datasets with dynamic labels).

• Example of training DyGFormer on Wikipedia dataset:

python train_node_classification.py --dataset_name wikipedia --model_name DyGFormer --patch_size 2 --max_input_sequence_length 64 --num_runs 5 --gpu 0

• If you want to use the best model configurations to train DyGFormer on Wikipedia dataset, run

python train_node_classification.py --dataset_name wikipedia --model_name DyGFormer --load_best_configs --num_runs 5 --gpu 0

• Example of evaluating DyGFormer on Wikipedia dataset:

python evaluate_node_classification.py --dataset_name wikipedia --model_name DyGFormer --patch_size 2 --max_input_sequence_length 64 --num_runs 5 --gpu 0

• If you want to use the best model configurations to evaluate DyGFormer on Wikipedia dataset, run

python evaluate_node_classification.py --dataset_name wikipedia --model_name DyGFormer --load_best_configs --num_runs 5 --gpu 0

We are grateful to the authors of TGAT, TGN, CAWN, EdgeBank, and GraphMixer for making their project codes publicly available.

Please consider citing our paper when using this project.

@article{yu2023towards,
  title={Towards Better Dynamic Graph Learning: New Architecture and Unified Library},
  author={Yu, Le and Sun, Leilei and Du, Bowen and Lv, Weifeng},
  journal={Advances in Neural Information Processing Systems},
  year={2023}
}