Thing is, a lot of these problems are always solvable my Tree (non DL) models if you design your data smartly. But GNNs have so much potential to capture information, especially with the use of context attention to get exactly the right patterns you're seeking.