Interesting idea. Thanks for sharing!

I only skimmed the article, so maybe they give a reason why they do this, but from "Algorithm 1" on the last page it looks like they extract features for each edge and then pass those features into the model. Many prominent AI researchers, like Andrew Ng, argue that you should perform end-to-end learning rather than feature engineering for better results.

Maybe that's only for high dimensional spaces and overkill for a simple low dimensional graph traversal algorithm? Feature extraction has the benefit of being more interpretable but usually at the expense of being suboptimal.

Cool to see Amboss working / publishing like this.