var.monotone option in GradientBoosting

[jingsliu07]

Hi, is it possible to add the equivalent of the var.monotone option in R GBM package to the GradientBoostingClassifier/Regressor? Sometimes it is really useful when we know/want some factors to have monotonic effect to avoid overfitting and non-intuitive results.

Thanks!

[gravesee]

+1 for this feature. It is what allows me to use GBM for regulated credit risk models

[tsterbak]

+1, I would love to use this!

[pat-oreilly]

Could someone possibly label this as a new feature and assign to me? I've already made a start: this notebook demonstrates it working for DecisionTreeRegressor and GradientBoostingRegressor. Code is here. I'd welcome advice on the API (and the code too), here's what I've gone with:

#Columns 0 and 4 monotonically increasing, column 2 monotonically decreasing
GradientBoostingRegressor(..., increasing=[0,4], decreasing=[2])

Thanks!

[amueller]

@PatrickOReilly please create a pull request so we can easily review the changes.
I'm not sure this is within scope, but it might be, depending on the changes. If we decide not to merge, it would still be great to have it in scikit-learn-contrib: https://github.com/scikit-learn-contrib/scikit-learn-contrib/blob/master/README.md