Regarding Probability Estimates Predicted by LIBSVM

Question

I am trying to classify 3 classes using the SVM classifier. How we interpret probability estimates predicted by LIBSVM. Is this based on the perpendicular distance of the specimen from the hyperplane of the maximum field ?.

Request, through some coverage, of the interpretation of the probabilistic estimates predicted by the LIBSVM classifier. The C and gamma parameters are first configured, and then probability estimates are output using the -b option for both training and testing.

Answer 1

Multi-class SVM always decomposes into several binary classifiers (usually it is a set of one and all classifiers). Any binary SVM classifier solution function outputs the (signed) distance to the separation hyperplane. In short, SVM maps the input region to a one-dimensional real number (solution value). The predicted label is determined by the sign of the value of the decision. The most common method for obtaining probabilistic output from SVM models is the so-called Platt scaling (document by LIBSVM) .

Is this based on the perpendicular distance of the specimen from the hyperplane of the maximum field?

Yes. Any classifier that displays such a one-dimensional real value can be subjected to further processing to obtain probabilities by calibrating the logistic function according to the values ​​of the classifier solution. This is the same approach as in standard logistic regression .

Answer 2

SVM performs binary classification. To achieve multiclass classification, libsvm does what it calls one against all. What you get when you call -b is the probability associated with this technique that you can find here here .