Gated gradient model reproducibility code

Run instructions

The repository contains the data and code needed to reproduce the experimental results of the paper "Learning node representations using stationary flow prediction on large payment and cash transaction networks". The bash script run_eth_experiment.sh will run the hyperparameter search for the gated gradient model and all the baselines using the ethereum data, which will have to be downloaded separately, see LINK. run_abl_unimodal.sh and run_abl_mulimodal.sh will reproduce the synthetic flow experimental results.

NB: Running all the experiments may take a long time. Using a GPU run_abl_unimodal.sh and run_abl_mulimodal.sh may take a few hours each and run_eth_experiment.sh may take up to 36 hours. To avoid having to rerun all experiments, the result files have been included in the results folder. The figures and tables of the paper can be reproduced from these results using the three notebooks.

Installation instructions

To install the minimal requirements run

pip install -r requirements_min.txt

This will install everything necessary to run run_eth_experiment.sh, run_abl_unimodal.sh and run_abl_mulimodal.sh.

• To run train_node2vec_emb.sh also install pytorch-geometric.
• To run preprocess_network_data.py also install graph-tool.
• To run the notebooks also install jupyter, matplotlib, seaborn and tikzplotlib.