no. For most models there is no natural way to handle missing data. For trees there are several "natural" ways. We try not to do too much magic under the hood, and so I think it makes most sense if the user selects the imputation algorithm themselves.

How would you support missing values in an SVM in a non-surprising way?

@amueller Thanks for the response. As I had pointed out in my original post, I was only referring to the class of algorithms where:

such support makes sense and is established in the literature, of course

For instance, you mentioned decision trees and I am myself working on refactoring the GMM so that it can be fit "naturally" without explicit imputation beforehand. Obviously not all models will have a "natural" approach that is established and it would be best to leave those alone, but for the ones that do (or will have) it might be something worth considering.

I'm not sure what this means.

Perhaps what it means is: test that meta-estimators support NaN. Perhaps the same is true of some transformers (?), affinity/distance-based learners, etc.

Yes, estimators, transformers, metrics, what have you. My point was, if there is a well established method of handling missing value for any of these algorithms then it might make sense to actually have them capable of doing so here in sklearn. Further, my post was geared more towards things that will be added in the future since it might be relatively less work for the original author compared to future refactoring by someone else. Of course, one can also consider the same type of NaN support for already existing algorithms, where feasible, however that would probably take a significant amount of work (is my mostly random guess).

Methods for handing missing values with SVMs do exist. See e.g. http://www.sciencedirect.com/science/article/pii/S0893608005001292 . However I don't know how good they are.

I vote for making all the generic enough algorithms like in preprocess to be usable in the presence of missing values. It's pain in the ass to reimplement preprocessing algos myself to preprocess the data for xgboost only because sklearn.preprocess is unsuitable because it is crippled and raises errors if it encounters nan.

In the meantime, we have moved towards adding native support for missing values in more and more of our estimators. So I guess we can conclude this issue.