This contains code for the experimental part of Expressive Power of Temporal Message Passing, in which the theoretical expressive power of various temporal message-passing formalisms are analysed.
The code relies on PyTorch and TGB.
First ensure that hyper-parameters in hyper.py are set correctly. Then, a link-prediction experiment can be run by invoking linkpred.py with two arguments: (i) the model type, either 'T1' or 'T2', corresponding to global and local models respectively and (ii) the TGB link-prediction benchmark you wish to use, e.g. tgbl-wiki.
Datasets are downloaded and processed automatically by the TGB infrastructure, with a prompt on first use. Coarse-grained status is provided on standard output, and fine-grained loss curves to TensorBoard runs/.