A GNN tutorial explains the principles clearly, but can't rank architectures
When you see a vendor claim that graph models can do both molecular screening and relation prediction, this tutorial explains the mechanism: information is passed back and forth between nodes, edges, and global attributes and then aggregated; the same model can make predictions at the graph, node, and edge levels, and shuffling node order leaves results unchanged.
Next time you see graph-model marketing, break it apart and ask: how many hops does information travel, at which layer is it aggregated, and does the prediction land on the whole graph, a node, or an edge — then talk about performance numbers.
If you want to compare which architecture is stronger, or cite results on a particular dataset, don't use this as your basis. It is only a survey, ran no new experiments, and the relative effectiveness of the various designs is unverified.
《A Gentle Introduction to Graph Neural Networks》(2021) | Next review 2027-09-20