Explaining GNNs with GNNExplainer
In this section, we will introduce our first XAI technique with GNNExplainer. We will use it to understand the predictions produced by a GIN model on the MUTAG dataset.
Introducing GNNExplainer
Introduced in 2019 by Ying et al. [2], GNNExplainer is a GNN architecture designed to explain predictions from another GNN model. With tabular data, we want to know which features are the most important to a prediction. However, this is not enough with graph data: we also need to know which nodes are the most influential. GNNExplainer generates explanations with these two components by providing a subgraph 
and a subset of node features 
. The following figure illustrates an explanation provided by GNNExplainer for a given node:
Figure 14.1 – Explanation for node 
’s label with 
in green and non-excluded node features 
To predict 
and 
, GNNExplainer implements an edge mask (to hide connections) and a feature mask...
