This post will likely get deleted. You should post in r/learnmachinelearning

Xgboost tends to be best on tabular data - you didn’t mention what domain you’re working on.

Creating entity embeddings from NNs and passing them to other downstream classifiers definitely is a thing