Research On Quality Of Experience And Behavior Of Mobile Network Based On Machine Learning

With the popularization of mobile networks and the maturity of network technology,the user's needs and business types are constantly enriched.The operators recognized that the most effective method to attract user and maintain user's loyalty is to really understand the user's Quality of Experience(QoE),and to improve QoE through adopting effective measures.Actually,QoE depends on the service(technology)factor,the surrounding factor and the user factor,it is a kind of service evaluation method with user's subjective recognition as cretirion.Currently,QoE research is mainly focused on the following two aspects:(1)the establishing of QoE model,including the definition,influencing factors,quantitative methods and evaluation methods of QoE.Due to the too many influencing factors towards QoE,referring to the knowledge and theory of computer science,psychology,sociology,statistics and so on.Therefore,the researchers usually evaluate and quantify the QoE focuzing on a specific mobile business from the technology level,resulting in the emerging of lot of different Qo E models established.(2)The exploring of user's behavior and regularity,it is beneficial to the improvement of QoE and the analysis of the differences of QoE satisfaction to investigate the rule of user behavior.In the era of mobile network with high quality and high technology level,user's QoE satisfaction is diverse from each other,the key reason is that the user's own behavior,the different degree of QoE satisfaction could be derived from different users.A clear understanding of the regularity of user behavior will be helpful for operators to optimize the network,allocate resources and provide personalized services for enhancing QoE.From the perspective of theoretical research and practical application,it is very important and challenge to study QoE model and user behavior rule.In this thesis,we aimed on constructing a comprehensive and efficient QoE model and exploring the regularity of user behavior,and conduted our researches with machine learning algorithm.By using the machine learning algorithm,we can deal with the problems of non-linear and super-high dimension of QoE influencing factors effectively,and find the complex and implicit user behavior rule.In addition,the operators can understand the real QoE satisfaction in the mobile network and carry out targeted operational strategies on account of the user behavior rule.The main contents and innovations of this paper are summarized as follows:(1)Based on the user-centered concept,the influencing factors of QoE in the mobile network environment are summarized from the aspects of technology,service and user.The mean opinion score(MOS)is used to quantify the QoE satisfaction,and the QoE is evaluated by the wavelet neural network based on improved step of glowworm swarm optimization(ISGSO-WNN)method.The QoE model with certain versatility for multi-mobile services is established.(2)In addition to the MOS method,user's complaint is another effective method of QoE quantification,which can truly reflect the satisfaction degree of QoE in mobile networks.In actual situations,even if the current satisfaction is low,most users have not complained,which belong to the implicit users.It is easy to form unbalanced problem in the data set.Based on the above facts,as a supplement to the QoE quantification and evaluation methods proposed in(1),and combined with the analysis results of the QoE influencing factors from the technical,service and user perspective,improved incremental fuzzy kernel regularized extreme learning machine(II-FKRELM)is proposed to evaluate QoE.By updating the output weights,using Cholesky decomposition and combination kernel functions,II-FKRELM can evaluate and predict the user's complaint in the mobile network quickly and accurately,and excavate the hidden complaint user effectively.Experimental results show that the proposed algorithm has advantages in unbalanced and super-high-dimensional data set.(3)User behavior has a profound impact on QoE,and different user behavior characteristics will lead to different user's satisfaction.At the same time,according to the analysis of the weight of the factors that affect the user's complaints,it also shows that the location and consumption behaviors have higher weights and greater impact on user's satisfaction.Based on the above analysis,and combined with the location information of user's complaint are scattered and disorderly in the mobile network,this paper focus on the cluster analysis of location information of complaint user,and digging out the trajectory rule of complaint user to take targeted measures to improve QoE for operators.Based on the analysis and summarization of the research results of trajectory clustering,adaptive density trajectory cluster(ADTC)on account of time and spatial distance calculation method is proposed.In ADTC algorithm,the cluster radius value is calculated from the density distribution of the data,the cluster center and the number of cluster are obtained by the corresponding adaptive strategy,and the new weighted rough C-means method is used to solve the boundary data problem.By the experiments on the public data set and the core network data set,the proposed ADTC algorithm based on time and space distance can convert the hard cluster into fuzzy cluster effectively,obtain the optimal cluster center and rich cluster results,and has the adaptive characteristics.Based on the above cluster results,operators can clearly understand the core area and scope of the mobile user trajectory,and take network optimization and resource adjustment to improve QoE.(4)The change of user consumption behavior is a clear reflection of the degree of satisfaction in mobile network.By studying the trend of user consumption behavior and understanding the current QoE satisfaction,it is beneficial to the operator to take measures to prevent the loss of user and save user.In this paper,fuzzy K-Prototypes cluster-based support vector machine(FKP-SVM)is proposed to solve the classification problem of mixed data in user consumption behavior,and improve the calculating speed of SVM in large sample data set.Through the comparison of the experimental results,the proposed new combined model method has a good performance in predicting the loss of user.