Can I remove the intercept?
DreHar opened this issue · comments
Hi All,
I'm sorry if I am just being dull, but is it possible to remove the intercept in GLM/GLMCV? I believe it is supported in the R version of glmnet (https://cran.r-project.org/web/packages/glmnet/ChangeLog version 1.9-3).
Could I comment out the intercept/beta0 calculations in pyglment.py and just set beta0 to 0?
Thanks!
It's not supported at the moment but it would be a welcome addition. You'd need a fit_intercept parameter which is a boolean
I have made an attempt to add in fit_intercept as a parameter and made a [WIP] PR - #310
Please let me know if there are any changes or adjustments. There are a few style points I wasn't sure about. I am not sure if it would be better to
- pass fit_intercept everywhere as a parameter
- change all methods to expect beta0 and beta and pass beta0 as 0 for the no intercept case everywhere
- in gradL2Loss, whether it should be one fit intercept check at the bottom and then calc by type, or multiple fit intercepts inside the current if else. Currently I just compute grad0 and then don't use it
What I have tried to do is add the fit_intercept parameter to the GLM and GLMCV classes. Then within all methods the state of fit_intercept is inferred. Then before finishing the GLM class will separate out the beta0_ and beta_. I had a look at sklearn.linear_models and this appeared to be how they handled it.
Sorry, I just realised I made a mistake and I do have some unit test failures. I will fix these up; any style or directional change still much appreciated
That seems looks fair and yeah, I would try to mimic sklearn as much as possible. Thanks a lot for the PR!
closed by #310