Classification score: SVM

Question

I am using libsvm for multi-class classification. How can I attach classification scores, to compare the confidence of classification, with the output for a given sample as:

Class 1: score1

Class 2: score2

Class 3: score3

Class 4: score4

Answer 1

0

SVM's are not a probabilistic tool, they are quite certain on their classification. There are certain instances when you use your results from a kernel machine to gets probs. Check the option -p and the use of support vector classification on SVM's, but actually for your output you should use something more intuitive, like Mixture of Gaussians, which are designed to give probabilities of the classes.

answered Jun 18 '11 at 22:00

Leon Palafox
3126●54●71●107

Which classifiers other than Mixture of Gaussians would give me the output as I want?

(Jun 19 '11 at 00:12) Xolve

1

Mhhh, as far as I understand, what you want is: p(class=class_k/new_sample)

which is excactly the kind of problem that mixture models solve and the exact output you wish. Youy basically are describing a clustering with probabilities.

You might try a multiclass logistic regression, which should give you something similar.

Remember, SVM's are pretty good, but there are problems other tools can solve better.

(Jun 19 '11 at 07:33) Leon Palafox

Answer 2

0	See these two references for a method that has worked pretty well for me in practice: Platt and Lin et al. I believe this method is implemented in libsvm. You could also just use logistic regression in LibLinear. answered Jun 19 '11 at 00:48 Lucas Wiman 136●1●5

Classification score: SVM

Subscription:

Related questions