Hi,

Thank you for this great article. Can you tell me the approach we should use to separate train/valid/test data sets for graph learning? I mean for an image classification problem we can separate 60% of images as train images, 20% as validation and the rest as testing. So, I feel that we need to separate graph nodes for a node classification/clustering tasks and we need to separate links(edges) of the graph for link prediction tasks? If I’m wrong please correct me. If I’m correct can you tell me the percentages we should consider for this separation?