Hi, Scott, I was also thinking about regression when I implemented the VotingClassifier but ultimately decided against it. I think that it could potentially be useful, but it is not as "generally applicable" as the VotingClassifier I'd say.
Implementation-wise, I think refactoring it into a VotingBase parent and a VotingClassifier and a VotingRegressor could technically be pretty straight-forward though. I am only a little concerned if averaging the results of different regressors could potentially help with the predictive performance. Would be interested to hear what others think. Has anyone averaged regression models in practice? > On a semi-related note, I notice that VotingClassifier.predict() doesn't > check to see if it has been fitted. Is there a general way to check whether > a classifier has been fitted yet? Good point! I will add a check for that. I think the general solution is the check_estimators_unfitted in from sklearn.utils.estimator_checks import check_estimators_unfitted Best, Sebastian > On Nov 10, 2015, at 1:09 PM, Scott Turner <srt19...@gmail.com> wrote: > > First of all, my thanks for all your hard work in providing Scikit-Learn. > It's a joy to use. > > Now that VotingClassifier has dropped, are there plans to create the > analogous ensemble classifier for regression, i.e., one that averages the > results of a list of base classifiers? I have an implementation that I would > be happy to submit if there's interest. (I'd suggest extending the > classifier to include a meta-classifier option as well as (say) 'mean' and > 'median'.) But maybe someone is already working on this? > > On a semi-related note, I notice that VotingClassifier.predict() doesn't > check to see if it has been fitted. Is there a general way to check whether > a classifier has been fitted yet? > > Finally, I've also implemented the CONFINE and CONVINE algorithms for > estimating the confidence of individual predictions from this paper > <http://journals.plos.org/plosone/article?id=10.1371/journal.pone.0048723>. > This isn't complicated -- it's a matter of doing k-nn on the training data > and then taking local error / variance means for the neighbors of the > prediction. It's nice in the sense that it can be applied to any estimator, > so it provides a way for getting confidence estimates (or at least local > error estimates) for estimators that don't provide that otherwise. If > there's interest I'm happy to contribute that code as well. > > -- Scott Turner > ------------------------------------------------------------------------------ > _______________________________________________ > Scikit-learn-general mailing list > Scikit-learn-general@lists.sourceforge.net > https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
------------------------------------------------------------------------------
_______________________________________________ Scikit-learn-general mailing list Scikit-learn-general@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
