Probabilistic version of the Support Vector Machines (SVM) model is the core of the Linear SVM model. It integrates a calibration approach based on Platt scaling to transform the output of the model into probability distributions. Since the calibration needs to be learned from unseen data, a cross-k-validation approach is adopted.
The probabilistic SVM model maintains the major properties of a linear SVM classifier, but it produces probabilities as prediction scores in a [0,1] range. Therefore, these values can be used as prediction confidences and they could enable post-processing steps that involve thresholds.