On GNN explainability with activation rules