Expressivity of Representation Learning on Continuous-Time Dynamic Graphs: An Information-Flow Centric Review

This repository contains the code supporting the work "Expressivity of Representation Learning on Continuous-Time Dynamic Graphs: An Information-Flow Centric Review".

Upon using this repository for your work, please cite this paper:

@article{ennadir2025if,
  title={Expressivity of Representation Learning on Continuous-Time Dynamic Graphs: An Information-Flow Centric Review},
  author={Ennadir, Sofiane and Zarzar Gandler, Gabriela and Cornell, Filip and Cao, Lele and Smirnov, Oleg and Wang, Tianze and Zólyomi, Levente and Brinne, Bj\"{o}rn and Asadi, Sahar},
  journal={arXiv preprint arXiv:2412.03783},
  year={2025}
}

Experiments

The experiments were run on a virtual machine with 32 vCPUs, default VM memory 128 GB and an NVIDIA L4 GPU with memory 24 GB GDDR6.

Long-range datasets

To reproduce results on the PascalVOC 10 and 30 datasets, use the original implementation here.

Bi-partite dataset

To reproduce results on the TGBL-Wiki dataset, create the environment by running bash create_env1.sh, and then run

cd src
bash run.sh

for DATA_NAME="tgbl-wiki" and a given MODEL - set on top of the bash script.

Community-based synthetic dataset

To reproduce results on the community-based synthetic dataset, create the environment by running bash create_env2.sh, and then create the synthetic data by running

cd synthetic_data
python3 sbm.py

Finally run

cd ../src/TGB
bash run_scripts/run_synthetic_data.sh

Homogeneous dataset

To reproduce results on the homogeneous dataset, create the environment by running bash create_env2.sh (if not already created), and then run

cd src/TGB
bash run_scripts/run_tgbl_coin_mini.sh

Tightness of theoretical bound

To evaluate the tightness of the theoretical bound, create the environment by running bash create_env2.sh (if not already created), and then run

cd src/TGB/bound
python3 test_bound.py --data tgbl-wiki --model-path <PATH_TO_MODEL_TO_EVALUATE>

by replacing <PATH_TO_MODEL_TO_EVALUATE> with the path to the .pth file.

Acknowledgements

This codebase is adapted from non-dissipative-propagation-CTDGs and TGB - the folder src/TGB includes a local, modified copy of this repository. We thank the non-dissipative-propagation-CTDGs and TGB authors for sharing their code. Please consider citing the authors from non-dissipative-propagation-CTDGs and TGB as well, in case this codebase is useful for your research.