An implementation of Latent-Graph Convolutional Networks based on PyTorch Geometric from the article:

Floris A.W. Hermsen, Peter Bloem, Fabian Jansen & Wolf B.W. Vos, End-to-End Learning from Complex Multigraphs with Latent-Graph Convolutional Networks. arXiv preprint arXiv:1908.05365, 2019 — arxiv.org.

See requirements.txt or environment.yml (conda).

The synthetic transaction networks can be found in the /data folder as zipped data files in pickle format. Testing can be done with files followed by a _tiny suffix. These contain less than 500 nodes and around 1250 transaction sets.

Model code can be found in model.py. Example notebook can be found in the main directory.