Requests for pull requests¶
We always welcome new feature requests and suggestions, but if you are looking to contribute, here are some features that we plan to add soon.
Exponential family distributions¶
One prominent next step is to increase the number of distributions available. In rough order of importance, we are looking at:
- Probit regression (see issue #159).
- Negative binomial regression (see issue #163).
- Other count distributions (e.g. Quasi Poisson, hurdle) (see issue #163).
- Cox regression
For each distribution, the typical workflow would involve
Cython implementation of coordinate descent¶
Coordinate descent requires nested for loops that can be sped up with cython. This would involve translating _gradhess_logloss_1d() and _cdfast() into cython code (see issue #104).
Provide GLMCV class for warm restarts¶
Currently, the GLM class returns a list of estimators corresponding to each value of \(\lambda\) using a warm restart approach. This not readily compatible with some of scikit-learn‘s cross-validation and grid search features.
This PR would require an overhaul of the existing GLM class in addition to writing a new class (see issue #158).