Network traffic classification plays an important role for network administrators to supervise the traffic flows in order to manage the network efficiently. Therefore, accurate classification of traffic flow is of great significance. The quality of traffic classification lies in the classifier’s accuracy and efficiency. This paper firstly implements an accurate and robust traffic classification solution using a recent new machine learning algorithm “Extreme Learning Machine”. This paper also proposes a way to automatically generating training dataset for traffic classifier, making the system adaptable and scalable for new network applications. In this paper, VoIP and WWW traffic are used as two categories for the traffic classification. Experiment results indicate that this solution is highly accurate, more stable and robust for the classification of our traffic flow samples compared with other methods such as C4.5, Random Forest, Na?ve Bayes and KNN sited in other literatures. The classifier proposed in this paper has 93% accuracy when test dataset collected after training dataset, and other algorithms have similar accuracy. When test dataset collected one and two months later than training dataset, it still keep 85% high accuracy while other algorithm only reach about 60% at most, whose accuracy deviation are too large to apply to the practical traffic classification system. Thus, it is quite accurate and robust, with great scalable for engineering practice.