Modified example of GNNExplainer usage with an adjusted coefficient for edge_size. It helps with stability and conciseness for explanations of small (10-100 edges) computation graphs

gnn_explainer_example.py

# Original example at 
# https://github.com/rusty1s/pytorch_geometric/blob/master/examples/gnn_explainer.py

# In this modified version we take into account datasets where many nodes have a
# 2-hop neighbourhood comprising 10 to 100 edges. In theses cases, the current fixed
# value of the edge_size coefficient is not working as intented. We propose a 
# variable coefficient, dependent on the number of edges of the 2-hop neighbourhood
# (we will consider this neighbourhood as the computation graph).
# More info at: https://github.com/rusty1s/pytorch_geometric/issues/1985
# or at @joaquincabezas 

import os.path as osp

import torch
import torch.nn.functional as F
import matplotlib.pyplot as plt
from torch_geometric.datasets import Planetoid
import torch_geometric.transforms as T
from torch_geometric.nn import GCNConv, GNNExplainer

# This import is needed for computing N (edges in the computation graph)
from torch_geometric.utils import k_hop_subgraph

dataset = 'Cora'
path = osp.join(osp.dirname(osp.realpath(__file__)), '..', 'data', 'Planetoid')
dataset = Planetoid(path, dataset, transform=T.NormalizeFeatures())
data = dataset[0]


class Net(torch.nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = GCNConv(dataset.num_features, 16)
        self.conv2 = GCNConv(16, dataset.num_classes)

    def forward(self, x, edge_index):
        x = F.relu(self.conv1(x, edge_index))
        x = F.dropout(x, training=self.training)
        x = self.conv2(x, edge_index)
        return F.log_softmax(x, dim=1)

# If we want to make reproducible experiments, remember to add here a fixed random seed
# seed = 1234
# torch.manual_seed(seed)
# random.seed(seed)

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = Net().to(device)
data = data.to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=0.01, weight_decay=5e-4)
x, edge_index = data.x, data.edge_index

for epoch in range(1, 201):
    model.train()
    optimizer.zero_grad()
    log_logits = model(x, edge_index)
    loss = F.nll_loss(log_logits[data.train_mask], data.y[data.train_mask])
    loss.backward()
    optimizer.step()

# We create the explainer before 
explainer = GNNExplainer(model, epochs=200, log=False)

# Node to be explained
node_idx = 10

# We obtain the computation graph of node_idx
_, _, _, hard_edge_mask = k_hop_subgraph(node_idx, 2, edge_index)

# N is the number of edges within the computation graph
N = torch.sum(hard_edge_mask==True).item()

S = 5  # Objective size of the explanation
L = 0.5  # Objective contribution to the loss function (for edge_size)

# Adjust the coefficient for edge_size:
# This formula is obtained considering that the value of a relevant edges,
# at the end of the process is around 0.9, while the value of irrelevant
# edges is around 0.1
# (https://github.com/rusty1s/pytorch_geometric/issues/1985)
explainer.coeffs['edge_size'] = 10*L/(N+8*S)

node_feat_mask, edge_mask = explainer.explain_node(node_idx, x, edge_index)
ax, G = explainer.visualize_subgraph(node_idx, edge_index, edge_mask, y=data.y)
plt.show()