Fast parameter optimization with randomized search

You might already be familiar with Scikit-learn's gridsearch functionalities. It is a great tool but when it comes to large files, it can ramp up training time enormously depending on the parameter space. For extreme random forests, we can speed up the computation time for parameter tuning using an alternative parameter search method named randomized search. Where common gridsearch taxes both CPU and memory by systematically testing all possible combinations of the hyperparameter settings, randomized search selects combinations of hyperparameters at random. This method can lead to a considerable computational speedup when the gridsearch is testing more than 30 combinations (for smaller search spaces, gridsearch is still competitive). The gain achievable is in the same order as we have seen when we switched from random forests to extremely randomized forests (think between a two to tenfold gain, depending on hardware specifications, hyperparameter...