Hello, I'm not sure I understand your problem. Of course, if you limit your dataset to class 2 and 4 to predict if your observation belongs to 2 or not (and 4 or not), this will perform better than if you have all the classes together. The latter is just much more complicated. If observations from classes 2 and 4 are much closer between each other than with observations in classes 1, 3, and 5, then maybe you should do a first classifier that predicts whether an observation x belongs to (2 or 4) or (1, 3 or 5). You can then use more complex model to distinguish between labels 2 and 4 (and 1, 3 and 5).
Cheers, Nelle On 9 June 2015 at 16:03, Herbert Schulz <hrbrt....@gmail.com> wrote: > Hello, > > My question is the following: > > i have an multiclass problem ( classes 1-5 prediction). > > If i split the data into classes 1,3,5 ( first dataset) and classes 2 and > 4 (second dataset). I trained models with every dataset separately. I'm > getting better results insteat of using all classes in one dataset and one > model. But the problem is, i dont know which model i should choose at the > beginning, cause i dont know if the input is in class 1,3,5 or 2,4. > > So in general: > > Trained model #1 with dataset containing classes 1,3 and 5 as targets. > Trained model #2 with dataset containing class 2 and 4 as targets. > > Is there a way to combine these models?? > > so depending on the Input, he should choose model #1 ore model #2 . but i > dont know how i should classifie them. > > Is there is a way with bayse or something like that? > > best regards, > > Herb > > > ------------------------------------------------------------------------------ > > _______________________________________________ > 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
