GraphLIME is a model-agnostic, local, and nonlinear explanation method for GNN in node classfication task. It uses Hilbert-Schmit Independence Criterion (HSIC) Lasso, which is a nonlinear interpretable model. More details can be seen in the paper.
This repo implements GraphLIME by using the awasome GNN library PyTorch Geometric, and reproduces the results of filtering useless features until now; that is, Figure 3 in the paper.
Please check that it is not the official implementation.
Just use pip to install.
> pip install graphlime
This implementation is easy to use. All you need to do is to confirm that your model outputs log probabilities (for example, outputs by F.log_softmax()) first, then instantiate a GraphLIME object, and finally explain the specific node by calling explain_node() method.
from graphlime import GraphLIME
data = ... # a `torch_geometric.data.Data` object
model = ... # any GNN model
node_idx = 0 # the specific node to be explained
explainer = GraphLIME(model, hop=2, rho=0.1)
coefs = explainer.explain_node(node_idx, data.x, data.edge_index)coefs are the coefficients of features. They correspond to the 𝜷 in the paper. The larger the value is, more important the corresponding feature is.
Example and details can be found in tutorial.ipynb.
All scripts of different experiments are in the scripts folder. You can reproduce the results by running the following command:
> sh scripts/noise_features_cora.sh
There are another 10 random features added to the original features and then we train a GNN model. We use several explanation methods to find out which features are important for the model. Then we plot the distribution of noise features selected by the explanation methods. Less noise features are better; that is, a peak around the origin is good.
- python >= 3.6
- torch >= 1.6.0
- torch-geometric 1.6.0
- numpy 1.17.2
- scikit-learn 0.21.3
- seaborn 0.10.1
- matplotlib 2.2.4
- 2021-02-13: add GNNExplainer in the noise features experiment
- 2021-02-05: add a tutorial file
- 2021-01-10: modified the file structure & published to PyPI with version 1.2.0
- 2020-11-26: fixed bug
- 2020-11-09: fixed bug
- 2020-10-08: modified the file structure for publishing to PyPI
- 2020-05-26: first uploaded