A Neuro-symbolic Approach to Enhance Interpretability of Graph Neural Network through the Integration of External Knowledge

Graph Neural Networks (GNNs) have shown remarkable performance in tackling complex tasks. However, interpreting the decision-making process of GNNs remains a challenge. To address the challenge, we explore representing the behaviour of a GNN in a representation space that is more transparent such as a knowledge graph, in a way that captures the behaviour of a GNN as a graph. Our initial experiments on the node classification task can represent the trained graph convolutional neural network (GCN) behaviour with some semantics uncovered by state-of-the-art approaches. This research offers a promising direction for enhancing GNN interpretability and understanding by providing structured, human-understandable representations and incorporating external knowledge for more accurate predictions.