How practical are logic engines for proof paths combined with knowledge graphs in Providing reasonable explainability for ML models trained using GNNs?
Adding more context. there is a history of logical reasoning (found in academic articles and in filed/granted patents) as a partial mechanism to building solutions to problems that are built on graphs. I am not an expert and thus I am not asserting that logic engines (like SWIProlog or any variant) are useful. I am curious about the use of ML models built on GNN (or GCN) methods where datasets are derived from larger and more complete knowledge graphs and any attempts to explaining how an observation is mapped to a prediction. This article (see sec. 4.1) cites work that uses proof paths to arrive at some approximation of explainability. I am looking for anecdotal examples of folks having worked with GNNs, KGs, and logic programming tools like prolog at showing how a prediction might be explained. If the cited work is sound, then this hints at logic engines being potentially useful for some use-cases.