Internet traffic classification and application identification associated with network traffic is an essential step for network security and traffic engineering. The traditional port-based, payload-based or statistic-based classification methods do not work well in practice. To solve the problem, in this paper, we propose a new classification model which is verified by supervised learning algorithms (e.g., RandomForest, C4.5, KNN). The model includes two main parts: a clustering flow label propagation technique based on equivalent flow-labeled propagation and a synthetic-flow feature generation algorithm based on Bidirectional-flow (BDF). To further improve the accuracy of traffic classifiers and reduce the cost, we also propose a feature selection algorithm. The experiments are done on real-world network traffic from a large-scale Supercomputing center, showing that the proposed model achieves better performance and higher accuracy for network traffic classification.