2013/2/15 Charles-Pierre Astolfi <[email protected]>: > On Fri, Feb 15, 2013 at 11:30 AM, Gael Varoquaux > <[email protected]> wrote: >> On Fri, Feb 15, 2013 at 11:22:02AM +0100, Charles-Pierre Astolfi wrote: >>> > However, before we do this, I'd like to understand: what is the usecase >>> > and the purpose for this function? >>> Which? lars_path or lasso_path or my proposition? >>> What I propose in an faster and almost (in the sense that it's not >>> computed via sgd and coefs may vary) drop-in replacement for >>> lasso_path. >> >> Small remark: lasso_path does a coordinnate descent (CD), and not an SGD. >> >> I understand that, but the whole point of lasso_path is to use a CD, and >> not a LARS, as the CD might be more stable, or even quicker in some >> situation. SO why do you want such a function? > It is quicker if one tries to do lasso one a small number of alphas. > But for a high number of alphas, I don't think it can ever be faster.
It probably depends on the shape of the data (n_samples, n_features) and the correlation between the features and the minimum value of alpha where you want to stop the path. >> I understand that. If you want to do parameter selection using Lars, why >> is LassoLarsCV not well-suited for what you want to do? I still don't >> understand what usecase you are trying to solve. > My setting was online learning with expert advice. At each time step, > you receive predictions from experts (models) for the next time step. > The goal is to find a linear combination of the predictions that beats > the best expert. > To find the linear combination I tried lasso and had to do parameter > selection at each step. CV is not used at all in this setting, mostly > because it (experimentally) does not achieve any significant gain. > Also, we now that the last time step is the one we want to optimize on > (since the undelrying distribution may vary wrt time). How do you evaluate which value of alpha is the best if you don't cross validate in one way or another? -- Olivier http://twitter.com/ogrisel - http://github.com/ogrisel ------------------------------------------------------------------------------ Free Next-Gen Firewall Hardware Offer Buy your Sophos next-gen firewall before the end March 2013 and get the hardware for free! Learn more. http://p.sf.net/sfu/sophos-d2d-feb _______________________________________________ Scikit-learn-general mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
