User Clustering And KQI Analysis In Cellular Networks

The drawbacks,such as passively problem detection and coarse-grained problem localization of traditional KPI(Key Performance Indicator)and customers' complaints based operation and maintenance modes are highlighted,especial for the purpose of CEM(customer experience management).Therefore,for cellular network providers,studying the methods for intelligent network operation and maintenance to explore the fine-grained CEM is an important research area for current and future.For the three steps of CEM lifecycle,intelligent network operation and maintenance should have the following abilities.1)Personalized preference analysis and prejudgment before the service.2)Detecting,delimitating or localizing experience problems in real time for dynamic network adjustment during the service.3)Optimizing network iteratively based on historical data after the service.In this dissertation,we focus on the modeling of personalized preference,detecting and delimitating experience problems.Due to the rapid development of big data technologies,cellular network providers gradually have the ability to the collect,store and operate huge amount of user data.Based on the which,the problems mentioned above have been studied widely.However,there still exist many challenges.For example,1)the huge volume of data makes the high complexity of CEM.2)The low quality of data makes user experience hard to characterize.3)The reasons which trigger bad experience are complex due to complex end to end link.In view of the above challenges,the specific contents of this dissertation includes:(1)Mobile user behavior clustering algorithm under massive and high dimensional user data is studied.In the context of user service usage behavior clustering,to solve the limitation of high time consumption of common used Kmeans clustering,this dissertation proposes AFKmc2+ clustering algorithm.AFKmc2+ first compress the original data through learning the data prototypes which is the idea of self-organizing mapping,and store the mapping relationships between original data and data prototypes.Then,based on the MCMC(Markov Chain Monte Carlo)sampling theory,the initial point selection step of Kmeans is done efficiently without loss much accuracy which has been proved.In the context of co-clustering user service usage behavior,to avoid three limitations of current tNMF co-clustering algorithm,namely,1)limitation of applicability due to high dimensions,2)number of clusters unknown before clustering and 3)hard co-clustering,this dissertation proposes H-tNMF co-clustering algorithm which,1)utilizing density clustering to construct scalable framework of clustering dimensions,in order to reduce multiply operations of big user matrix,2)utilizes the idea of hierarchical clustering to determine the number of clusters automatically,3)dimensions sharing mechanism for child co-clusters based on the definition of Consistent Facto,which is utilized to evaluate the co-cluster's quality,to prevent error accumulation introduced by hierarchical clustering.Utilizing open data set and our user temporal-spatial traffic usage data set,the performance improvement of AFKmc2+ and H-tNMF are verified.(2)The freeze identification algorithm of HTTP video service is studied from network side without user feedback.To solve the limitations of current freeze Identification algorithm,namely,1)narrow range,2)low efficiency and 3)low accuracy.Through analyzing the phenomenon of "time drifting" which induce low identification accuracy,this dissertation proposes black box strategy based freeze Identification method.Instead of focusing on how to rebuild terminal video from network side,the proposed strategy only concerns the relationship between TS features at network side and video freeze at terminal side.For feature extraction,this dissertation proposes Freeze Divine algorithm.Based on freeze mechanism,Freeze Divine estimates the remaining playing time in current buffer based on the partial poor network state characterized by TS parameters defined in xDR equipment.Overall,the combination of black box strategy and Freeze Divine algorithm translate the traditional "rebuild and hard-decision" freeze Identification strategy to "estimation and soft-decision".Utilizing our video freeze data set constructed by our KQI-doctor platform,the performance improvement of proposed method is verified.(3)The method of anomaly root cause delimitation from massive Service Provide RTT(spRTT)is studied in the context of complex service usage scenarios.To avoid the model complexity of scenario-based spRTT induced by scenarios differences,this dissertation proposes a Divide and Conquer based Context Modeling(DC-CoMo)algorithm,DC-CoMo utilizes robust multiple linear regression method to,1)avoid the impact of outliers,2)consider the comprehensive impact of scenarios to spRTT.DC-CoMo first utilizes hierarchical dimension aggregation method to avoid the sparsity of scenario-based spRTT,then utilizes the idea of divide and conquer to do trade-offs between model complexity and model accuracy.To avoid the limitation of current anomaly root cause delimitation methods which cannot delimitate anomaly,namely,1)induced at small scale,2)induced by complex triggers,3)multiple independent triggers,at the same time.Based on the anomaly detected by DC-CoMo,this dissertation proposes a Kullback-Leibler divergence based greedy search tree(ReasonTree)algorithm.ReasonTree first utilizes Kullback-Leibler divergence to quantify the degree of distinction between scenario-factors to anomaly,which make the algorithm sensitive to root cause happens in small scale.Through the construction of search tree,different scenario-factors are combined to quantify the degree of distinction between their combinations to anomaly,thus,ReasonTree can delimitate anomaly induced by complex triggers(multiple scenario-factors).Based on the greedy strategy,once the most influential root cause is found,the corresponding scenario-based spRTTs will be dropped and another delimitation step goes on until no anomaly exists,thus multiple independent root causes can be delimitated.ReasonTree does not rely on history delimitations,thus has ability to find new ones.Utilizing the spRTT data set collected from xDR equipment in 4G cellular network,the performance improvement of proposed DC-CoMo and ReasonTree are verified.Based on the big data captured by our self-developed xDR-Pro and KQI-Doctor platform from real cellular network,this dissertation focuses on the problem of personalized preference modeling,network side user experience anomaly detection and delimitation with the usage and improvement of existing clustering,classification and anomaly detection algorithms.The proposed methods and strategies provide new ideas for the fine-grained CEM for intelligent operation and maintenance of cellular network.