Analysis Of User Traffic Behavior In Wireless Networks Based On Machine Learning

With the rapid development of mobile communication networks,the number of mobile devices has surged,a wider variety of services appears,and the user's service quality is increasingly improving.The accurate portrayal of the user's network traffic behavior can help to improve the customized service for the user and provide useful information for the network optimization.Individual user tends to have large divergence in their behaviors.Due to more serious fluctuation in user's network traffic,existing time-series prediction algorithms cannot be directly applied to the network traffic prediction of individual user.In this thesis,the user network traffic prediction problem is analyzed and studied.The main contributions of the thesis include:User traffic time domain distribution and user behavior analysis is performed.After preprocessing and aggregating the raw data,the user traffic data time series is obtained,and the periodic and auto-correlation characteristics of the overall network traffic are analyzed.Then for individual user,the characteristics of user traffic usage,the mobile phone application usage and user mobility pattern are analyzed.Results show that there exists a significant long tail effect in the user's traffic behavior,and the user mobility has a great impact on the user's data traffic and the user's application usage behavior.High-mobility mobile users tend to use more traffic and access more mobile apps.Clustering of user network traffic data is conducted.The explanatory factor analysis method is used to extract the feature of the user's traffic data time series and reduce the dimensionality.Then,five common factors are extracted,which represent five specific time periods respectively.The K-means clustering algorithm is used to cluster the users.And based on the time domain characteristics of the user traffic data,the similarity between users is discovered.The users are finally divided into six types.This provides the possibility of userspecific charging policy based on the prediction result of user data traffic.The K-means clustering based on factor analysis method has better performance and more interpretable features.Based on the clustering of user mobile application usage,the correlation between user application behavior and user's traffic data is analyzed.Latent semantic analysis is applied to extract the feature and reduce the dimensionality,then K-means clustering is adopted to cluster the user into six different types.Based on analyzing of the traffic data of different types of users,it is found that the behavior of users using traffic data is closely related to their application usage behavior.Therefore,the mobile application usage is helpful for predicting the user traffic.User traffic prediction algorithm is investigated.A user network traffic prediction model based on Prophet algorithm and Gaussian process regression model is proposed.Wavelet transform is first employed to decompose the original user traffic data time series into low frequency component and high frequency component.The low frequency component bears the long range dependence of user network traffic,while the high frequency component reveals the gusty and irregular fluctuations of user network traffic.Then Prophet algorithm and Gaussian process regression algorithm are used to predict the low frequency component and high frequency component respectively.Using the mean absolute percentage error as the evaluation metric,compared with the existing time series prediction algorithm,the proposed scheme can lower the prediction error by half.