Hi Michael,

thanks for your answer, it helped a lot already.

On 06.10.2014 11:06, Michael Eickenberg wrote:
> may prefer using e.g. LassoLarsCV.

I gave it a shot and indeed the coefs are available, and it worked 
really fine on my well designed test problems. However, with my real use 
case it performed poorly in comparison to LassoCV when compared to 
unseen data because of stability issues.

Is there an intrinsic problem for LassoCV not being able to return the 
coefs?

> Highly correlated designs generally result in a strong support
> identification instability of these algorithms. It depends on what you
> care about to be able to conclude if the warnings you see are relevant
> or not. If it is prediction and your variables are highly correlated,
> then whichever is selected, they will predict similarly.

I am indeed interested in prediction and that is why I wanted to include 
as many predictors as possible. I removed all predictor pairs correlated 
to more than 0.98 (completely arbitrary) but I think this is the wrong 
strategy now. I'll filter some correlated predictors beforehand and try 
again with both LassoCV and LassoLarsCV

Thanks a lot!

Fabien





------------------------------------------------------------------------------
Slashdot TV.  Videos for Nerds.  Stuff that Matters.
http://pubads.g.doubleclick.net/gampad/clk?id=160591471&iu=/4140/ostg.clktrk
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general

Reply via email to