Feature Selection And Classification Of Internet Video Traffic

With the development of Internet and streaming media technologies, the proportion of video traffic in the network traffic is continuously increasing. Internet video traffic classification allows ISP(Internet Service Provider) to allocate the Internet resource according to the different levels of QoS(Quality of Service) requirement of various Internet video applications, which can help improve the network planning and resource utilization.This paper studies seven typical video applications, including three kinds of Web video with different definitions, QQ video, network live video, P2 P client video and Http downloads video, and proposes a method for Internet video classification in a finer granularity. The main work is as follows:The paper presents CON-GR(Consistency Feature Selection-Gain Ratio) feature selection method, which considering both the consistency and relevance. Firstly, the method quickly removes the useless QoS related features by CON and then ranks the remaining features in descending order based on information gain ratio. Finally, the optimal feature subset is chosen on the basis of the classifier. Compared with three other feature selection methods, CON-GR can effectively reduce the training time and computational complexity by decreasing feature dimensions while ensuring equivalent accuracy. In addition, this paper analyzes the differences of statistical features among Web video with three definitions, as well as the differences between Web video and network live video.The modified two-layer SVM(Support Vector Machine) classifier is designed to classify the Internet video traffic in a finer granularity. The first layer is M parallel binary classifiers, and the second layer is an M-class classifier. The first layer makes the classifier better than others, because we can select effective features separately for each video application, which only need to distinguish one application with all others, instead of distinguish all video applications. Moreover, the input of the second layer is the CON-GR feature selection results for all video applications, instead of the feature collection of the first layer. About 10% of the flow data was sent to the second layer for further classification, which proves the necessity of the modified two-layer classifier for fine-grained classification. Experiment results show that the proposed method can improve significantly the overall accuracy for the Internet video traffic classification in all scenarios. We also verify that the proposed method is promising for short-time traffic identification.