2
$\begingroup$

I used gaussian process classification implemented in Matlab gpml toolbox and also in R kernlab. For my problem - 600*14 matrix with two classes - it trains order of magnitude slower than other classifiers. Gauss. process is taking hours to finish what SVM/naive bayes/random forest does within minutes.

Is this expected behavior of the classifier, or is there some other reason for such a poor speed? (e.g. scale of the data, slow implementation, etc.)

$\endgroup$
1
  • $\begingroup$ naive GP classification is $O(N^3)$ where $N$ is the number of samples... $\endgroup$
    – Memming
    Aug 3, 2015 at 13:52

1 Answer 1

1
$\begingroup$

GPML generally tunes the hyper-parameters as well as fitting the model, which takes quite a while. This ought to be done for SVMs as well for a like-for-like comparison (SVMs can take a very long time to train depending on the hyper-parameter values). If you are using the minimize function, that probably explains the difference.

$\endgroup$

Your Answer

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

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