Explaining GNNs via iterated decision trees.

This is the code accompanying the paper

@article{ldgnn,
  author  = {Alexander Pluska and Pascal Welke and Thomas Gärtner and Sagar Malhotra},
  title   = {Logical Distillation of Graph Neural Networks},
  year    = {2024},
  comment = {under review},
}

Setup

We recommed using a conda or virtual environment with pip installed. The required packages can then be installed by calling

pip install .

in the project directory. This will likely install a CPU version of pytorch. In order to use GPU acceleration, please consult the PyTorch documentation on how to install the correct version for your system.

In order to run the experiments in evaluate.sh/evaluate.py you also need to set up wandb.

Running the experiments

For convenice, we provide a jupyter notebook example.ipynb that demonstrates training a GCN and an IDT on the AIDS dataset with a single random split. It should run fine on most devices and configurations.

The experiments in the paper are performed using parallel 10-fold cross-validation and can be run via

.\evaluate.sh

They are configured for a node with 4 GPUs. If you have fewer GPUs, you can adjust the --devices argument in the evaluate.sh script. CPU is not supported.