Companion code for the paper "New Perspectives on the Evaluation of Link Prediction Algorithms for Dynamic Graphs"

tgn_attn contains a simple training script for the TGN-Attn model.

The code used to generate the plots in the paper is available in the notebooks folder as jupyter notebook.

Getting Started

Create a virtual environment

python3 -m venv venv
source venv/bin/activate

Install the requirements

pip install -r requirements.txt

Model Weights

In the examples plots provided, pre-trained models are used. I can provide the weights for the models upon request. Else the scores and weights can be obtained using:

• The tgn_attn script for the score against selected negative samples experiment
• By adapting the code in https://github.com/shenyangHuang/TGB/blob/main/examples/linkproppred/tgbl-wiki/tgn.py

For any questions, please contact raphael.romero@ugent.be.