I'm using libsvm (executables provided in the package.) for a 2-class classification problem. In training step I'm setting b parameter as 1, which is defined as;
-b probability_estimates : whether to train a SVC or SVR model for
probability estimates, 0 or 1 (default 0)
and in predicting step I'm setting b parameter as 1 again, which is defined as;
-b probability_estimates: whether to predict probability estimates, 0 or 1 (default 0); for one-class SVM only 0 is supported
But based on the formulation of SVM, prediction results are signed distances to the margin. My question is, when trained SVM predicts results as probability estimates, how does it (libsvm) map the distances to probabilty estimates?
Thank you
解决方案
They use an improved version of Platt's method.
You can find here:
http://www.csie.ntu.edu.tw/~cjlin/papers/libsvm.pdf (Section 8, probability estimates)
and here is the description of the method:
http://www.csie.ntu.edu.tw/~cjlin/papers/plattprob.pdf