Thanks for the links Kacper. Explainability is another important topic, which is however not specific to GNNs but a general issue in deep learning. I find that in some settings learning the graph itself (rather than having it given) is a way towards (some) explainability. I will try to write a post about it.