yngvem / group-lasso

Group Lasso implementation following the scikit-learn API
MIT License
105 stars 32 forks source link

Poisson regression #6

Open yngvem opened 4 years ago

yngvem commented 4 years ago

There has been several requests for group lasso regularised Poisson regression. This was more difficult than first thought as the gradient of the Poisson loss is not globally Lipschitz.

The loss gradient is, however, locally Lipschitz. A line-search for the Lipschtiz constant can therefore converge (and will under some reasonable assumptions). It is unclear if the restarting scheme will work with a line-search so that need to be investigated further.

Unfortunately, the code is currently structured so that major refactoring is needed to support a line-search for the Lipschitz.

yngvem commented 2 years ago

The package is now using a line search, so this is possible to implement, but I don't currently have the time for that.