Figure 2:

A. GNN computation graph G_c (green and orange) for making prediction $\widehat{y}$ at node v. Some edges in G_c form important neural message-passing pathways (green), which allow useful node information to be propagated across G_c and aggregated at v for prediction, while other edges do not (orange). However, GNN needs to aggregate important as well as unimportant messages to form a prediction at node v, which can dilute the signal accumulated from v’s neighborhood. The goal of GnnExplainer is to identify a small set of important features and pathways (green) that are crucial for prediction. B. In addition to G_S (green), GnnExplainer identifies what feature dimensions of G_S’s nodes are important for prediction by learning a node feature mask.