2
$\begingroup$

I read in several places that SVM is a discriminative model, but SVM has no statistical aspects per se, by that I mean that is does not estimate any probablity, specifically the postirior distribution as a discriminative model should do.

Shouldn't SVM be considered as nirther disciminative nor genreative, rather it finds a disciminative hyperplane in the feature space.

$\endgroup$
1
  • $\begingroup$ Not all discriminative models estimate the posterior distribution, some just estimate decision boundaries between classes, like SVM and decision trees. Discriminative models just enable "discriminating" between classes. $\endgroup$
    – noe
    Dec 30, 2021 at 15:46

1 Answer 1

1
$\begingroup$

Support Vector Machines are discriminative because they fit a hyperplane which separates two classes. So it learns a decision boundary which is the definition of discriminative methods.

$\endgroup$
2
  • $\begingroup$ Well, according to Bishop's PRML (p 43 microsoft.com/en-us/research/uploads/prod/2006/01/… ) "Approaches that model the posterior probabilities directly are called discriminative models.", this is why I am confused $\endgroup$
    – Benny K
    Jan 8, 2022 at 22:26
  • $\begingroup$ According to Jebara, Tony (2004). Machine Learning: Discriminative and Generative: Even if a classification does not use any probability it is still discriminative. But still: Franc et al. (icml.cc/Conferences/2011/papers/386_icmlpaper.pdf) can show that SVM are a Probabilistic Model with conditional probabilities which meets your definition of Bishop and the general one as well. $\endgroup$
    – technik
    Jan 9, 2022 at 8:08

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

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