A probabilistic support vector machine for uncertain data

A probabilistic support vector machine (PSVM) is proposed for classification of data with uncertainties. Performance of the traditional SVM algorithm is very sensitive to uncertainties. The noises in input space will cause uncertainties of the mapping in feature space. The traditional SVM algorithm may not be effective when uncertainty is large. A new probabilistic optimization is proposed to determine the decision boundary. The minimal distance is described probabilistically by its probability distribution function. Finally an artificial dataset and a real life dataset from UCI machine learning database are used to demonstrate the effectiveness of the proposed PSVM.