Abstract: Compared to support vector machine (SVM), relevant vector machine(RVM) is much more appropriate for online classification, and provides probabilistic classification. Based on the research and analysis of probabilistic classification and its influence on overall accuracy, a new online traffic classification method is proposed. Firstly, the method utilize RVM to classify traffic flows and output probabilistic classification. Then, re-identify the flows, whose classification probability is in douting interval [0.1, 0.9], by using port & DPI. And otherwise, absolutely accept the classifcation when predicted probality is in the interval [0, 0.1] and [0.9, 1]. Experiment studies illustrate that the method can reach the overall accuracy of 98%, and performs well in online network traffic classification.

Key words: traffic classification, relevant vector machine, traffic features, douting interval