I just wrote a Kaggle kernel showing how to use the knnExtract() method from my R package fastknn to make a nonlinear mapping from the orginal features and improve H2O GLM performance. The resulting method is very fast and performs better than Random Forests for the Forest Cover Type dataset.

Take a look at it to learn how to use fastknn to improve your performance on Kaggle competitions, and let me know what you think!