A Gentle Introduction to Graph Neural Networks

Ended up using these quite a bit during my thesis, they are great for certain types of data. I read a lot of papers claiming these could a solution to the causality problem. Hence why you often see papers using GNNs named something something causal something. The Technical implementation however always felt a bit lackluster. The passing algorithms are at best 1 extra pass, but often were implemented as an extra function per link, scaling them quadratically per node. This was always a bit sad as it prevented large node GNNs. Havent looked at them the last years, might need to do a new pass myself to see if they have evolved a bit. The article is quite good at introducing it, but does indeed miss the performance problem here a bit :)