search

stanfordmlgroup / ngboost

Natural Gradient Boosting for Probabilistic Prediction
Apache License 2.0
1.62k stars 214 forks source link

Mixture Density Networks with ngboost #134

Open astrogilda opened 4 years ago

astrogilda commented 4 years ago

Hey @alejandroschuler! Is it possible to create an MDN with ngboost? That is combine multiple distributions (in the most general case, from different families) to predict the output of interest?

alejandroschuler commented 4 years ago

Yes, in theory, but you would need to create a new distribution class (e.g. ThreeNormalMixture, which would have three mean parameters and three variance parameters). It would be cool to have a class factory mixture(Dist, K) that generates the appropriate distribution class on-the-fly. Honestly not sure off the top of my head if that's possible or easy.

JoGall commented 8 months ago

Did either of you ever experiment with implementing this @astrogilda or @alejandroschuler? And if not, did you ever find another framework capable of probabilistic forecasting with multimodal output distributions?