Figure 1:

GnnExplainer provides interpretable explanations for predictions made by any GNN model on any graph-based machine learning task. Shown is a hypothetical node classification task where a GNN model Φ is trained on a social interaction graph to predict future sport activities. Given a trained GNN Φ and a prediction ${\widehat{y}}_i=“\mathrm{Basketball}”$ for person v_i, GnnExplainer generates an explanation by identifying a small subgraph of the input graph together with a small subset of node features (shown on the right) that are most influential for ${\widehat{y}}_i$. Examining explanation for ${\widehat{y}}_i$, we see that many friends in one part of v_i’s social circle enjoy ball games, and so the GNN predicts that v_i will like basketball. Similarly, examining explanation for ${\widehat{y}}_j$, we see that v_j ‘s friends and friends of his friends enjoy water and beach sports, and so the GNN predicts ${\widehat{y}}_j=“\mathrm{Sailing}.”$