移动通信网中的用户聚类与KQI分析

发布时间：2018-04-28 10:22

本文选题：用户聚类 + 关键质量指标　；参考：《中国科学技术大学》2017年博士论文

【摘要】：传统的基于关键性能指标(KPI:Key Performance Indicator)和用户投诉的体验运维方式已暴露诸多弊端,如:发现问题具有被动性、局限性,解决问题精细化程度不足等。对于电信运营商来说,探索面向智能化运维的精细化客户体验管理是当前及未来的重要研究领域。面向客户体验生命周期中体验前、中、后三大环节,智能化的客户体验管理应具备如下能力:在业务发生前进行个性化用户偏好需求分析与预判,制定事前预防策略;在业务发生过程中快速发现体验问题、定位/定界问题并动态实时调整;在业务结束后基于历史数据学习进一步迭代优化网络,最终形成智能运维的闭环。本文的研究问题聚焦在体验前用户的个性化偏好行为建模、体验中的准确监测体验问题和定界问题的相关策略。得益于大数据技术的快速发展,电信运营商逐步具备了海量用户数据采集、存储和计算的条件和能力,上述问题得以广泛研究,但仍存在诸多挑战,如:数据规模巨大,体验管理复杂度极高;数据质量不高,体验信息难以准确刻画;体验异常成因复杂。针对上述挑战,本文具体研究内容包括:(1)移动用户的业务行为偏好建模,该研究针对大规模用户的精细化体验管理面临的实现复杂度超高的问题,提出了基于于聚类的用户行为偏好建模方式,平衡精细化的度和体验管理的复杂度。针对Kmeans聚类方法存在的聚类速度慢问题,提出的AFKmc2+两阶段聚类算法:首先利用自组织映射思想将原始数据映射为数据原型,然后借助马尔科夫蒙特卡洛(MCMC:Markov Chain Monte Carlo)采样理论快速、合理选择原型初始聚类中心,最终利用Kmeans完成快速聚类。针对基于双侧正交约束的非负矩阵分解(tNMF:Tri-Non-negative Matrix Factorization)的双向聚类算法存在聚类个数事先未知、频繁大矩阵相乘及硬聚类策略导致聚类效果差问题,提出的H-tNMF算法:借助密度聚类思想实现聚类维度可伸缩,减少频繁大矩阵相乘操作;引入分裂层次聚类思想自动确定聚类个数;定义双向簇紧致度概念,实现子簇维度共享软聚类,避免错误累积。利用公开数据集以及现网用户数据集验证所提算法性能,实验结果证明:相比于AFKmc2算法,AFKmc2+在具有更快的聚类速度的同时具有相接近的聚类精度;相比于tNMF算法,H-tNMF在相同聚类数目下具有更好的聚类效果,且支持软聚类。(2)网络侧视频点播业务卡顿识别,该研究针对采集条件制约导致的视频传输状态数据质量不高问题,提出基于特征构造结合黑箱策略的卡顿建模方法,方法克服数据中存在的偏差,可以准确识别出视频卡顿、长卡顿和多次卡顿。通过分析导致现有卡顿识别方法准确率低的"时间漂移"现象,提出一种不进行视频重建,而只关注如何构建网络侧视频数据流和终端侧视频卡顿间的映射关系的黑箱策略。通过剖析视频卡顿机理,提出一种基于网络状态局部极差值的缓冲区剩余数据量估计方法(Freeze Divine)用于映射关系学习。所提方法将现有卡顿识别方案中的"重构+硬判决"的识别思路转变为"估计+数据驱动下的软判决"。通过构造现网视频卡顿数据集验证所提算法性能,实验结果证明:所提方法的卡顿识别准确率比现有方法高20%;Freeze Divine所提去的卡顿相关特征较其他特征与视频卡顿的相关性更强。(3)网络侧 HTTP 业务响应延迟(spRTT:Service Provider Round Trip Time)异常根因定界,该研究针对端到端业务体验异常成因不单一问题,提出基于多元线性回归的分治场景建模方法(DC-CoMo:Divide and Conquer based Context Modeling)和基于相对摘的贪婪搜索树构建方法(ReasonTree),高效标定异常,准确定界异常根因。提出DC-CoMo算法:利用层次维度聚合思想降低spRTT时序稀疏性;利用分治思想进行海量spRTT时序建模,权衡spRTT场景模型的复杂度和准确度;建模采用多元线性回归方法,综合考虑异常点及场景对spRTT的综合影响。针对现有业务KQI异常定界方法难以同时定界小范围异常根因、多维异常根因和多类型并存的异常根因问题,提出ReasonTree算法:算法首先利用相对熵量化场景属性对异常的区分度,定界小范围异常;通过构建搜索树的方式计算不同场景属性组合的异常得分,定界多维异常;利用贪婪迭代策略,识别异常主导原因并过滤对应数据,定界多类型并存异常;ReasonTree不基于历史定界结果,能够发现新的异常根因。利用现网HTTP业务spRTT数据集及人工注入异常数据集验证所提算法性能,实验结果表明:对于人工注入的三种异常根因,所提ReasonTree方法均具有超过95%的异常根因定界准确率;结合ReasonTree算法,DC-CoMo算法利用少量模型便可定界传统逐一建模算法约93%的异常根因。本文以自主研发的xDR-Pro和KQI-Doctor平台为基础,以现网多点实测的用户业务数据为支撑,通过对现有聚类、分类和异常检测等算法的使用和改进,研究了面向智能运维的精细化客户体验管理中的移动用户偏好建模、网络侧视频点播业务卡顿识别以及SP响应延迟异常定界问题。所提方法和策略为面向智能运维的精细化客户体验管理提供新思路。
[Abstract]:The traditional way of experience and maintenance based on KPI:Key Performance Indicator and user complaint has exposed many disadvantages, such as: finding the problem with passivity, limitation, and the lack of fine resolution. For the telecom operators, the exploration of the refined customer experience management for the intelligent operation and maintenance is the current and An important research field in the future. In the three major links of experience, in the customer experience life cycle, and in the post experience, the intelligent customer experience management should have the following ability: to analyze and prejudge the needs of personalized user preferences before the occurrence of business, and to formulate pre event prevention strategies; to find the experience problem quickly in the process of business affairs, and to locate / delimit the boundary. The problem and dynamic real-time adjustment; after the end of the business, further iterative optimization network based on historical data learning, and finally the closed loop of intelligent operation and maintenance is formed. The research problem of this paper focuses on the personalized preference behavior modeling of the user before experience, and the related strategies of accurate monitoring experience and delimiting problems in the experience. The rapid development of the telecom operators has gradually acquired the conditions and capabilities of mass user data collection, storage and computing. The above problems have been widely studied, but there are still many challenges, such as: the large scale of the data, the high complexity of experience management, the poor quality of the data, the difficult to accurately depict the experience information, and the complicated causes of the experience of the abnormity. The specific research contents of this paper include: (1) the business behavior preference modeling of mobile users. This study aims at the complexity of the sophisticated experience management of large-scale users, and proposes a clustering based user behavior preference modeling method, balancing the degree of refinement and the complexity of experience management. The AFKmc2+ two stage clustering algorithm proposed by the class method is slow. First, the original data is mapped to the data prototype using the self organizing mapping idea, and then the initial cluster center of the prototype is selected reasonably with the help of Markoff Montecarlo (MCMC:Markov Chain Monte Carlo) sampling theory. Finally, the Kmeans is used to complete the rapid completion of the cluster. For the bidirectional clustering algorithm based on tNMF:Tri-Non-negative Matrix Factorization based on bilateral orthogonal constraints, the number of clustering numbers is unknown in advance, the frequent large matrix multiplication and the hard clustering strategy lead to the poor clustering effect. The H-tNMF algorithm is proposed to achieve the scalability of the clustering dimension with the aid of the density clustering idea. Reduce the multiplicative operation of frequent large matrix, introduce the split hierarchical clustering idea to determine the number of clustering automatically, define the concept of two-way cluster tightness, realize the subcluster dimension sharing soft clustering, avoid error accumulation. The performance of the proposed algorithm is verified by the open data set and the current network user data set. The experimental results show that AFKmc2+ is compared to the AFKmc2 algorithm. Compared with tNMF algorithm, H-tNMF has better clustering effect under the same number of clustering and supports soft clustering compared with tNMF algorithm. (2) video on demand network side video on demand service card recognition, which is a problem of low quality of video transmission state data caused by acquisition conditions, proposed Based on the Caton modeling method which combines the feature construction with the black box strategy, the method overcomes the deviation in the data, and can accurately identify the video Caton, the long Caton and the multiple Caton. By analyzing the "time drift" phenomenon which leads to the low accuracy of the existing Caton recognition method, a kind of non video reconstruction is proposed, but the network side is only concerned about how to build the network side. The black box strategy of the mapping relationship between video data stream and terminal side video Caton. By analyzing the video Caton mechanism, a buffer residual data amount estimation method (Freeze Divine) based on the local difference value of the network state is proposed for mapping relationship learning. The proposed method identifies the "reconstruction + hard decision" in the existing Caton recognition scheme. Do not change to the "soft decision of estimated + data driven". By constructing the present network video card data set to verify the performance of the proposed algorithm, the experimental results show that the accuracy rate of the proposed method is 20% higher than the existing method; the correlation feature proposed by Freeze Divine is more relevant than its feature and video carton. (3) network The exception root of the side HTTP service response delay (spRTT:Service Provider Round Trip Time) is bound. This study aims at the unitary cause of the abnormality of the end to end business experience, and proposes a multi linear regression based modeling method for the divide and conquer scenario (DC-CoMo:Divide and Conquer based Context Modeling) and the greedy search tree based on the relative plucking. Construction method (ReasonTree), high efficiency calibration anomaly, accurate bound anomaly root cause. Propose DC-CoMo algorithm: use hierarchical dimension aggregation idea to reduce spRTT time series sparsity; use divide and conquer idea to model mass spRTT time series, weigh the complexity and accuracy of spRTT scene model; modeling using multiple linear regression method, comprehensive consideration of anomaly points And the comprehensive effect of the scene on spRTT. Aiming at the problem that the existing business KQI exception bound method is difficult to be fixed at the same time, the ReasonTree algorithm is proposed. The algorithm first uses the relative entropy to quantify the diversity of the anomaly area by using the relative entropy to quantify the anomaly area. The way of cable tree is used to calculate the anomaly score of the combination of different scene attributes, and the boundary multidimensional anomaly. The greedy iteration strategy is used to identify the abnormal leading causes and to filter the corresponding data, and the bound is multiple types of abnormality. ReasonTree can discover new abnormal root causes without the result of historical delimiting. It can use the spRTT data set and artificial injection of the current network HTTP service. The outlier dataset validates the performance of the proposed algorithm. The experimental results show that, for the three abnormal root causes of artificial injection, the proposed ReasonTree method has more than 95% anomaly root due to the demarcation accuracy, and the DC-CoMo algorithm uses a small number of models to delimit about 93% of the abnormal root cause of the traditional one by one modeling algorithm. Based on the developed xDR-Pro and KQI-Doctor platform, based on the user service data of the current network, the existing clustering, classification and anomaly detection algorithms are used and improved. The mobile users are well modeled in the refined customer experience management and the network side video on demand service card recognition is studied. And the SP response delay delimitation problem. The proposed method and strategy provide a new idea for the refinement of customer experience management for intelligent operation and maintenance.

【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TN929.5

【参考文献】