Explainable GNN-based models over knowledge graphs

Graph Neural Networks (GNNs) are often used to realise learnable transformations of graph data. While effective in practice, GNNs make predictions via numeric manipulations in an embedding space, so their output cannot be easily explained symbolically. In this paper, we propose a new family of GNN-based transformations of graph data that can be trained effectively, but where all predictions can be explained symbolically as logical inferences in Datalog—a well-known knowledge representation formalism. Specifically, we show how to encode an input knowledge graph into a graph with numeric feature vectors, process this graph using a GNN, and decode the result into an output knowledge graph. We use a new class of \emph{monotonic} GNNs (MGNNs) to ensure that this process is equivalent to a round of application of a set of Datalog rules. We also show that, given an arbitrary MGNN, we can extract automatically a set of rules that completely characterises the transformation. We evaluate our approach by applying it to classification tasks in knowledge graph completion