Feature Selection Toward Optimizing Internet Traffic Behavior Identification

摘要

P2P and multimedia similar applications are seemed as primary bandwidth consume network behaviors. Accurate network traffic behavior identification supports numerous network activities from network management, monitoring and Quality-of-Service(QoS), to forecast and application-specific investigations. Accuracy and performance are the two most important metrics for traffic identification especially for online implementation. In this paper, the optimization of feature selection to traffic identification is demonstrated in two traces which are captured from different time and location. Moreover, this optimization to traffic identification toward various applications are compared and analyzed in online and offline status with C4.5 decision tree algorithm. Our research demonstrated that the optimal features for traffic identification are mainly sensitive to application, time and location. Identifying for the same application behavior on different network location are sensitive to different features. Experiment result shows that the selected optimal feature subset can greatly improve the performance for both online and offline identification. Furthermore, it can improve the online traffic identification implementability in real network condition.