Is there a way to convert discriminant values in SVMLight multi-class classification into probability scores

Question

I am using SVM Light multi-class classifier for training a classifier with four classes. In the classification stage the classifier outputs the predicted label and the scores for the 4 classes. As the SVM Light website says, these scores are "the discriminant values for each of the k classes". I want to show the probability value of each of the class to the users. So I was wondering if there is some mathematical trick or some other way using which I can "convert" these values into probability values or at least into a normalised score in between 0 and 1?

Answer 1

SVM won't lead to probability estimates. There are some tricks for doing so, but they require additonal models, like for example Platt's scaling.

On the other hand, if you just want to "scale" the arbitrary number to [0,1] interval you can use the sigmoid function to do so.