Skip to content

Latest commit

 

History

History
78 lines (62 loc) · 4.62 KB

README.md

File metadata and controls

78 lines (62 loc) · 4.62 KB

L2XGNN: Learning to Explain Graph Neural Networks

The repository contains the code to reproduce the results of our paper. The source code consists of two folders:

  • benchmark_gc to reproduce the experiments related to graph classification tasks.
  • xai_evaluation to run, train, and evaluate L2XGNN in comparison with common XAI post-hoc techniques.

Requirements

The scripts are implemented with Python 3.9, and tested with Linux OS.

  • pyg==2.0.3
  • pytorch==1.10.1
  • networkx==2.6.3
  • numpy==1.21.2
  • dive-into-graphs==0.2.0
  • scikit-learn==1.0.2
  • matplotlib=3.5.1

Graph Classification

The folder benchmark_gc contains the evaluation script for various methods on common benchmark datasets via 10-fold cross-validation, where a training fold is randomly sampled to serve as a validation set. We slightly modify the evaluation protocol to work for our case.

Base model hyperparameter selection

Before training L2XGNN, we need to find the best configuration for each base model considered (i.e., GCN, GIN and GraphSAGE). Hyperparameter selection is performed for the number of hidden units [16, 32, 64, 128] and the number of layers [1, 2, 3, 4] with respect to the validation set. First, run the following command:

  • python 1_backbone_selection.py

L2XGNN hyperparameter selection and evaluation

Once the hyperparameters of the default models are found, we select the best ratio R from the set of values [0.4, 0.5, 0.6, 0.7] based on the validation accuracy.

  • python 2_ratio_selection.py --connected={} - connected: parameter to decide between connected and disconnected subgraphs (default value True).

Finally, we can select the perturbation intensity lambda_ from the list of values [10, 100, 1000]. This parameter can be manually changed in the following file l2xgnn/graph_imle.py. Below, the corresponding code snippet for the connected strategy:

class GraphIMLETopKConnected(torch.autograd.Function):
    tau: float = 1.0
    lambda_: float = 10.0    # choose from [10.0, 100.0, 1000.0]
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

Once the perturbation intensity is decided, we can run the following command:

  • python 3_main.py --connected={}

Explanation Evaluation

The folder xai_evaluation contains the scripts to train, evaluate, and plot the explanations obtained with our XAI method on a 3-layer GIN architecture.

Datasets

The folder datasets contains the raw data used for our experiments: ba_2motifs and Mutagenicity (MUTAG_0). The first one can be directly obtained using commands from common libraries. The latter was manually downloaded from this repository. The subfolder data_splits contains the splits used for this experiment.

Before training L2XGNN, we need to preprocess the MUTAG_0 dataset using the following command:

  • python 0_preprocess_Mutagenicity_data

Train L2XGNN

To train L2XGNN on a 3-layer GIN architecture (as in the paper), use the following command:

  • python 1_l2xgnn_train.py --dataset={} --model={} --connected={} --ratio={} --split={}
    • dataset: choose between ba_2motifs and Mutagenicity.
    • model: base GNN model we want to explain. Architectures supported: GIN, GCN, and GraphSAGE. You can choose between L2XGIN, L2XGCN, and L2XGSG respectively (default L2XGIN).
    • connected: parameter to decide between connected and disconnected explanatory subgraphs (default value False).
    • ratio: ratio of restrained edges (float between 0.1 and 0.9).
    • split: data split to evaluate (integer in the range [0,4]).

Evaluate L2XGNN

Once the explanations are obtained, we can compute the explanation accuracy in comparison with the available ground-truth motifs:

  • python 2_l2xgnn_evaluate.py --dataset={} --model={} --connected={} --ratio={} --split={}

Plot explanations

Finally, by running the next command, we can generate and save the images of the explanations learned by our method.

  • python 3_plot_explanations.py --dataset={} --model={} --connected={} --ratio={} --split={}

Reference

@article{serra2024l2xgnn,
  title={l2xgnn: learning to explain graph neural networks},
  author={Serra, Giuseppe and Niepert, Mathias},
  journal={Machine Learning},
  pages={1--23},
  year={2024},
  publisher={Springer}
}