implement categorical distribution for multiclass classification

[alejandroschuler]

https://github.com/stanfordmlgroup/ngboost/commit/87386b69f9eeba9e0f65c4afc4e8e7b187df0c2c

Notes:

• To make this work with the existing code (which assumes each distribution has a fixed number of parameters), the categorical distribution with K classes is generated by a class factory function so the model is instantiated like: ngb = NGBClassifier(Dist=k_categorical(3))
• The array of outcomes Y should be an integer array with values from 0 to K indicating the number of the class that Y[i] takes, as is standard in sklearn.

still to do:

• fix the predict method for NGBClassifier to work for multiclass problems (https://github.com/stanfordmlgroup/ngboost/commit/eb3cd4efb447affbb4fa1f0b6ea1ac0cb6931722)
• implement CRPS score for categorical distribution
• write an example script

[alejandroschuler]

note: may also want to change the distributions accepting the transpose of the params... makes the code annoying

[chemadix]

if my k classes were defined from a number different from "0" (e.g. from 1 to 10); how can the paremeter Dist can be modified? Is it necesary to use k_categorical?

[alejandroschuler]

@chemadix it is a requirement that the classes be 0,1,2... K-1 for Dist=k_categorical(K). You have to transform your Y so they are numbered that way before running ngboost. In the case you describe, just take Y=Y-1.