scikit-learn-contrib / lightning

Large-scale linear classification, regression and ranking in Python
https://contrib.scikit-learn.org/lightning/
1.73k stars 214 forks source link

[MRG] correct step size. #53

Closed fabianp closed 8 years ago

fabianp commented 8 years ago

The value of alpha needs to be multiplied by n_samples to match the formulation of objective.

Also, small fix for when the input is sparse but not in CSR format

fabianp commented 8 years ago

BTW I didn't manage to make a test for this one as I can't reproduce it on synthetic data. I can however see that it improves convergence on my own dataset.

fabianp commented 8 years ago

@casotto confirmed to me privately that this fixed his convergence issues too. changing the subject to MRG

mblondel commented 8 years ago

I will check the formulas on my side. Give me a few more days.

fabianp commented 8 years ago

Hmm, I tried to redo the math but I think I was wrong and they are currently correct. I'll need to dig further into this, closing it for now.