I've mostly come to the conclusion that DL is not going to beat tree based models especially when it comes to training speed. What I am now looking at is if it

I've mostly come to the conclusion that DL is not going to beat tree based models especially when it comes to training speed. What I am now looking at is if it's possible to feed in world knowledge from models like beer as features to XGBoost but without resorting to creating templates for table rows