Highest Probability SVM Nearest Neighbor Classifier for Spam Filtering

In this paper we evaluate the performance of the highest probability SVM nearest neighbor classifier, which is a combination of the SVM and k-NN classifiers, on a corpus of email messages. To classify a sample the algorithm performs the following actions: for each k in a predefined set {k1, ..., kN} it trains an SVM model on k nearest labelled samples, and uses this model to classify the given sample, then fits a sigmoid approximation of the probabilistic output for the SVM model, and computes the probabilities of the positive and the negative answers; than it selects that of the 2 × N resulting answers which has the highest probability. The experimental evaluation shows, that this algorithm is able to achieve higher accuracy than the pure SVM classifier at least in the case of equal error costs.