Evaluation of Support Vector Machines and Minimax Probability Machines for Weather Prediction

The paper evaluates two kernel-based methods on the problem of predicting precipitation based on observable variables. The support vector machine (SVM) method finds the two parallel hyperplanes that provide maximal separation of two subsets, excepting outliers. The minimax probability machine (MPM) method finds an optimal separating hyperplane that minimizes the probability of misclassification. Both SVM and MPM are binary classification methods that can be extended easily to multiclass problems. Both make use of the “kernel trick” to transform a linearly inseparable problem into a higher dimensional space where the problem may be linearly separable. The paper also investigates the accuracy and Peirce Skill Score (PSS, also known as the Hanssen and Kuipers discriminant) measures resulting from adding derived variables and removing variables. Using cross validation on the training data the accuracy was 70% and PSS 47%. On the final testing data the PSS was 35%.