1
$\begingroup$

I have been working on online learning for a few weeks now, especially with Vowpal Wabbit and logistic regression. My understanding of the online learning algorithms and the problem is alright but I can't get my head straight about the regularisation issue.

In a standard machine learning problem, one uses a validation dataset in order to tune regularisations parameters for L1 and L2. But how do you choose those regularisation parameters in an online setting? Do you just fix them from start, or should I update them while training occurs?

$\endgroup$
2
  • 1
    $\begingroup$ Do you train your model with the full dataset every time you get new data? $\endgroup$ Apr 24, 2018 at 13:53
  • $\begingroup$ No, only the new data is fed to the model $\endgroup$
    – Alexis
    Apr 24, 2018 at 14:03

1 Answer 1

1
$\begingroup$

As you say @Alexis, cross-validation is the standard way of selecting a model (or its parameters) in offline learning (especially when previous knowledge in not available). For online training, one usually incorporates some sort of (automatic) compensation for the model parameters that adjust them at the same time that training takes place (the alternative being to cross-validate on some gather data or equivalently to pre-optimize your parameters). Check also this answer as well as, for example, this paper (it shows that this issue is a very active research area)!

$\endgroup$
2
  • $\begingroup$ If I understand correctly you can use standard regularisations methods such as L1 and L2 by performing a beforehand training on available data. In an online setting there are regularisation strategies such as truncated gradients or per feature learning rate but there is no standard method for updating L1 and L2 parameters in an online setting for instance? $\endgroup$
    – Alexis
    Apr 25, 2018 at 15:50
  • $\begingroup$ Yes, I understand there are various approaches to online model selection in general. For example. see the "rolling origin", common in the context of forecasting, which is in effect an "application" of cross-validation idea to time-series: otexts.org/fpp/2/5 $\endgroup$
    – Sotiris
    Apr 26, 2018 at 14:43

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.