Abstract: The dynamic characteristics of burstiness and self-similarity make it difficult to classify the various applications from internet traffic. By analyzing the imbalance and heavy-tailed distribution of internet traffic, a heavy-tailed traffic based classification model for quantitative analysis is proposed. Both the sampling position and the window of training set in the long tail of traffic are studied. The SVM algorithm is adopted for the traffic classification, since the minor one included in a combined traffic is generally small sampling. The experimental results show that the heavy-tailed traffic can obtain a better classification mode by choosing best sampling position and scale of training set. Moreover, the quantitative analysis model has positive significance for achieving better classification precisions.