Feature request: Multi-target, feature subset ensembling, and active memory limits

[ldmtwo]

I love what this can do and it appears better than any other ensembling libs I've found so far, but there are a few things I would really like to see, either in commercial only H2O or in the open source, free for all.

If there is any further development, here are the features I would like to see (if you don't mind):

• Allow for multiple targets: Sometimes, we have multiple unknowns that are related to our primary target. I've found it helpful at work to model each target and use the probabilities/estimation of each target to make a final prediction on the primary target. As an example, imagine that the Titanic problem was to detect survived or not, with gender being another dependent variable and all the other variables left as independent.
• Feature subsets of the independent variables (X): I think it speaks well for it's self. Ex: only using numeric log transformed variables in Lasso, but all variables in RF.
• Have a limit to the number of active models in memory; a way to repeat a model with different seeds; and a way to persist/load and freeze models in lower levels so that the top level model can be tuned.

Thanks!