Multitask Learning For Class-of-Service Network Traffic Classification

In the practice of network flow classification,network operators usually only need to know the Class of Service(CoS)required by network flow to make decisions on network flow priority and resource allocation.In order to meet the needs of users for experience quality,a service-oriented multi task classification method of network flow is adopted;This method directly classifies cos oriented streams without inferring application types.In this paper,a multi task learning framework for service level is proposed.The macro feature group is defined by domain knowledge,the Shapley value model and Pearson correlation coefficient in cooperative game are applied to reasonably screen the features,the selected features are mapped into cos tasks,and the decision tree is used to solve the problem of cos threshold division.Experiments are carried out with real network data sets.In the case of a small amount of labeled data,the network parameters are optimized and the stable correlation coefficients of time loss and classification accuracy of each network model are adjusted.The results show that the classification accuracy and time consumption of this method are better than the existing literature methods.At the same time,the experimental results of multi classification are analyzed and some suggestions are given.The main contributions of this thesis are as follows:(1)Screening out suitable feature maps as CoS tasks,including classifying features into macro features based on domain knowledge,performing Shapley value(SV)analysis on each type of macro features,according to the contribution of the feature set in the interpretable model,The subset of features that contribute the most to the model algorithm is selected from all feature sets.At the same time,the Pearson Correlation Coefficient between the feature subsets is analyzed,and the feature that contributes the most to the Multi-task Model for Class-of-Service Network Traffic Classification(MTM-CoS)is selected,which can not only reduce the time consumption of the model algorithm,which also increased the interpretability of the model.(2)The screened features are further mapped to CoS tasks.Different from the existing methods that select CoS tasks from the perspective of users and the difficulty of acquiring features,the method in this thesis considers the rationality of selecting CoS tasks.The distribution of each CoS task is analyzed,and the decision tree binning(DTBin)algorithm is used to optimize the division of each CoS threshold(CThre)to improve the overall accuracy.Existing methods divide the threshold through histogram and linear calculation,which easily leads to an increase in the misclassification rate in the preprocessing part.In this thesis,the rationalization of threshold division by DTBin can effectively reduce the misclassification rate.(3)This thesis conducts experiments on the single-task Convolutional Neural Networks(CNN),Transfer Learning(TL)and MTM-CoS on the ISCX and Data set of Nanjing University of Posts and Telecommunications(ND)respectively,and compares them with existing literature.Degree and time loss performance comparison.The experimental results show that MTM-CoS inherits the advantages of deep learning to automatically extract features,and can achieve good classification accuracy with a small amount of labeled data.At the same time,the network parameters are optimized,and the stable correlation coefficient between the time loss and classification accuracy of each network model is adjusted.