I'm comparing the results of an Newton optimizer for a modified version of SVM ( a generalized quadratic loss, similar to the one stated in:

A generalized quadratic loss for SVM

) with classic SVM^light for regression. The problem is that it's able to overfit the data (UCI Yacht data-set) but I can't reach the generalization results of SVM^light. I've tried several hyper-parameters grids. I'm solving the primal problem. I'll send you my code if you need it. Any suggestion?

  • $\begingroup$ I figured out I've not shuffled the whole data-set (training+validation+test). Now it generalizes in a better way (at least on the UCI Yacht data-set) $\endgroup$ – filippo portera Nov 26 '19 at 12:56
  • $\begingroup$ I was wrong. I tried a different shuffle and the problem still persist. $\endgroup$ – Filippo Portera Dec 17 '19 at 10:33

Your Answer

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

Browse other questions tagged or ask your own question.