I applied Platt Scaling after a linear SVM, then applied the model on the training data just to see how the probabilities look like. I expect that the column "prediction(class)" is equal to class "positive" whenever "confidence(positive)" > 0.5. But for a considerable number of feature vectors, this was not the case, as if the threshold is not 0.5.
I suspected that the reason could be that those feature vectors are in the margin of the svm, so I used RBF kernel with very large C and got the results as expected (i.e. 0.5 is the threshold).
thanks in advance