We saw that CART methods trained with ensemble routines are powerful when it comes to predictive accuracy. However, they can be computationally expensive and we have covered some techniques in speeding them up in sklearn's applications. We noticed that using extreme randomized forests tuned with randomized search could speed up by tenfold when used properly. For GBM, however, there is no parallelization implemented in sklearn, and this is exactly where XGBoost comes in.
XGBoost comes with an effective parallelized boosting algorithm speeding up the algorithm nicely. When we use larger files (>100k training examples), there is an out-of-core method that makes sure we don't overload our main memory while training models.
The biggest gains in speed and memory can be found with H2O; we saw powerful tuning capabilities together with an impressive training speed.
