I support the inclusion of Poisson loss, although a quick note on predict_prob_at:
The output of Poisson regression is a posterior distribution over the rate parameter in the form of a Gamma distribution. If we assume no uncertainty at all in the prediction, the posterior predictive distribution (different from posterior distribution) is a Poisson distribution, otherwise it is a Negative-Binomial distribution with the extra parameter due to the variance of the prediction. If the prediction uncertainty can’t be computed, I’d suggest not including a predict_proba_at method at all, due to the overconfidence it would constantly imply. If a zero-variance posterior approximation is used, that should be definitely be noted somewhere. This problem doesn’t arise in logistic regression for example because the class probabilities are independent of the variance of the posterior distribution. More info: http://www.markirwin.net/stat220/Lecture/Lecture4.pdf Also, statsmodels implements poisson loss and could be a good reference. Brian Scannell On Thu, Jul 30, 2015 at 11:12 AM Mathieu Blondel <math...@mblondel.org> wrote: > > > On Thu, Jul 30, 2015 at 11:38 PM, Andreas Mueller <t3k...@gmail.com> > wrote: > >> I am mostly concerned about API explosion. >> I take your point of PDF vs PMF. >> Maybe predict_proba(X, y) is better. >> Would you also support predict_proba(X, y) for classifiers (which would >> be predict_proba(X)[np.arange(len(y)), y]) ? >> > > That could indeed be a good idea for consistency between classifiers and > regressors. > For Poisson regression, if y is not passed, i.e., if predict_proba(X, > y=None) is used, we could possibly use np.unique(y_train) by default. > This is obviously not possible for GPs, though. > > ------------------------------------------------------------------------------ > _______________________________________________ > 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
