We consider the explanation problem of Graph Neural networks (GNNs). Most existing GNN explanation methods identify the most important edges or substructures for deterministic graph. However, many systems represented by graphs in real life are inherently uncertain, and data collection is in itself an imperfect process, resulting in connections having varying likelihoods of existence. This work proposes a novel method, to explain the prediction of edge probability on uncertain knowledge graph by GNN. My supervisor is Ece Calikus, affiliation is Uppsala University.