Graduate to GNNs: Graph embeddings are “shallow,” which limits the amount of complexity they can handle and the utility for out-of-sample predictions. GNNs extend embeddings, encoding them more deeply with graph and node structural feature-based information included. They are the “turtles all the way down”6 stage of graph data science maturity. Graphs are used as input, predictions are then made on the graph, and the resultant output is in the form of a graph.