1. Thesis section

Original paper

Good paper interpretation

For node classification tasks, GNNExplainer realizes: extracts the subgraph structure and some features most relevant to the current node prediction result, and uses mask to indicate
How to find it: By continuously adjusting the subgraph G_s and featuresSubset X_s (which feature dimensions are working, not which node features are working), complete maximization Mutual information task, calculation formula:

Y is the prediction result is fixed, the function H represents entropy, so H(Y) is constant, so the goal is to minimize the above formulaSubtraction:

The above formula is based on The calculation of conditional entropy has been expanded. I have been confused for a long time before, and the paper has come to this point with a snap

2. Framework provided by pyg

from torch_geometric.nn import GNNExplainerimport matplotlib.pyplot as pltexplainer = GNNExplainer(model, epochs=300) # model isNode to be explainednode_idx = 3840 # index of the node to be explainednode_feat_mask, edge_mask = explainer.explain_node(node_idx, data.x, data.edge_index) # x/edge_index corresponds to the nodeFeature/Edge Setax, G = explainer.visualize_subgraph(node_idx, data.edge_index, edge_mask, y=y)# y is the prediction resultplt.savefig("explain_" + str(node_idx) + ".png")plt.show()