1、硕 士 学 位 论 文基于用户中心点访问上下文的边缘缓存应用研究专业名称 ： 电路与系统基于用户中心点访问上下文的边缘缓存应用研究I摘要移动用户激增和流媒体业务兴起造成了回程链路带宽的巨大消耗，新型 5G结合移动边缘计算（Mobile Edge Computing，MEC ）架构将内容投递移动到网络边缘，做到请求本地产生本地消费，能有效减少回程链路消耗。另一方面，由于边缘缓存容量的限制，高效智能的边缘缓存策略也成为了该领域的研究热点。边缘缓存的一个特点在于内容请求受用户时空移动性的影响，同时内容的时域局部性也造成了边缘侧内容请求的多样性，因此学习用户时空上下文和探索内容请求多样性能潜在提高边缘
2、缓存策略缓存效率。本文研究内容如下：1针对边缘侧内容请求受用户时空移动性的影响和多样性的特点，本文采用上下文多臂问题对边缘缓存问题进行建模。由于内容流行度的时变性和求解目标的 NP 难特性，本文提出基于用户中心点访问上下文的在线边缘缓存策略（Online Edge Caching Strategy based on User Central Point Access Context，OCUC ），并详细阐述了 OCUC 的三个主要模块：用户特征构造模块、群体上下文构造模块和结合上下文的在线探索模块。2利用边缘侧用户访问时空特性存在的中心点效应，本文提出用户中心点特征转换算法，将传统的用户时空特
3、征转换成用户中心点特征，并利用距离相关系数（Distance Correlation，DC）分析用户中心点特征的有效性。结果表明，用户中心点特征与用户访问兴趣具有更高的相关性。3内容缓存受用户群体兴趣偏好影响，针对用户群体中用户个数不定和群体上下文有效性评价的问题，本文结合用户特征，提出基于卷积神经网络的群体偏好上下文构建方法（CNN for Group Preference Context，GCNN），并进一步通过实验分析了网络超参数对网络输出的影响。4考虑到真实场景中上下文空间为连续空间和内容空间大造成的过量探索问题，本文结合多臂算法提出上下文缩放缓存算法，对受欢迎的上下文区域进行缩放，从
4、而维护更细粒度的上下文与内容偏好关系。5基于真实的中国移动用户记录数据集，本文从缓存击中率、算法稳定性和运行效率 3 个方面对不同的边缘缓存更新算法进行对比。结果表明，OCUC在缓存击中率和算法稳定性上均优于所有的对比算法；在运行耗时上，OCUC略多于 MPC 算法，所以实际应用中不会造成系统过载。本文提出的 OCUC 策略能有效学习用户时空上下文与内容偏好关系，实现基站间经验共享，从而有效减轻回程链路带宽消耗。本文研究工作不仅对用户时空上下文相关的边缘缓存策略设计具有参考意义，也具有一定的应用价值。硕士学位论文II关键词：边缘缓存，时空上下文，卷积神经网络，上下文缩放，多臂算法基于用户中心点
5、访问上下文的边缘缓存应用研究IIIAbstractThe proliferation of mobile users and the rise of streaming media services have resulted in huge consumption of backhaul link bandwidth. To effectively solve this problem, the new 5G network architecture combining with mobile edge computing (MEC) is proposed to move content
6、delivery to the mobile edge, so that requests can be served locally. Meanwhile, due to the limitation of the edge cache capacity, efficient and intelligent edge cache strategy has become a hot topic in this field.One characteristic of edge cache is that the content request is sensitive to the spatio
7、-temporal mobility of the user. At the same time, the temporal locality of the content also results in the diversity of content requests on the mobile edge. Therefore, the study of spatio-temporal context of the user and the exploration of the diversity can potentially increase the efficiency of edg
8、e cache policy. The main achievements of this paper are listed as follows:1. Content request on the mobile edge is affected by the users spatio-temporal mobility and shows diversity, Thus this paper uses the contextual multi-armed bandit problem to model the edge cache problem. Due to the time-varia
9、tion of content popularity and the NP-hardness of the solution, we proposes an online edge caching strategy based on user central point access context (OCUC) and elaborates on OCUCs three main modules: user central point feature construction module, group context construction module, and context-spe
10、cified online exploration module.2. Based on the central point effect in the spatio-temporal characteristics of user on the mobile edge, this paper proposes a central point feature transform algorithm to convert the traditional user spatio-temporal features into user central point feature. Moreover,
11、 we uses distance-correlation coefficients (DC) to analyze the correlation between user features and user interests. The result shows that the correlation between central point feature and interests is higher.3. The content cache is influenced by the user groups interest preference. To solve the pro
12、blem of uncertainty of user population and effectiveness of group context evaluation on user groups preference context construction, this paper proposes a group preference context construction method (GCNN) based on convolutional neural network. Furthermore, the influence of the number of user centr
13、al point, the number of convolution kernels, and the output dimension of the full-connected layer is analyzed.硕士学位论文IV4. Considering the problem that the context space is a continuous space and the over-exploration caused by the large content space in the real scene, this paper proposes a contextual
14、 zooming algorithm (CZ), which zooms in the popular context area to maintain finer-grained relationship between the context and the content preference.5. Based on the real China Mobile User Detail Record (UDR), this paper compares OCUC and other edge cache update algorithms from cache hit rate, algo
15、rithm stability and operating efficiency. The results show that OCUC is superior to all comparison algorithms in terms of cache hit rate and algorithm stability. In terms of operating efficiency, OCUC takes slightly more time than MPC, which means OCUC will not cause system overload when applying to
16、 real scenes .The OCUC strategy proposed in this paper can effectively learn the relationship between the users spatio-temporal context and content preference, realize the sharing of experience among the base stations, and thus effectively reduce the backhaul link bandwidth consumption. The research
17、 in this paper can not only provide a new insight for the design of edge caching strategy, but also has practical value.Keywords：Edge Cache, Spatio-temporal Context, Convolutional Neural Network, Context Zooming, Multi-armed Bandit基于用户中心点访问上下文的边缘缓存应用研究V目录摘要 .IAbstract .III1 绪论 .11.1 研究背景及意义 .11.2 国内
18、外研究现状 .31.2.1 边缘缓存策略研究现状 .31.2.2 用户访问时空特性研究现状 .51.3 本文的主要内容及结构安排 .62 边缘缓存问题建模 .72.1 问题描述与建模 .72.2 系统框架 .82.3 本章小结 .103 用户中心点特征构造 .113.1 用户访问时空特性 .113.1.1 用户访问可预测性 .113.1.2 时间特性 .123.1.3 空间特性 .133.2 用户中心点特征构造方法 .143.2.1 时空中心点定义与提取 .153.2.2 中心点特征转换算法 .163.2.3 中心点特征有效性分析 .173.3 本章小结 .194 用户相关的群体偏好上下文构造
19、 .204.1 方法概述 .204.2 基于 CNN 的群体偏好上下文构建方法 .214.2.1 CNN 原理 .214.2.2 GCNN 网络结构 .264.2.3 网络超参数对 GCNN 输出影响 .274.3 群体偏好上下文构建方法对比 .314.4 本章小结 .335 结合上下文的在线探索缓存算法 .345.1 方法概述 .345.1.1 上下文无关的探索算法 .345.1.2 上下文相关的多臂算法 .365.2 上下文缩放缓存算法 .375.3 本章小结 .406 实验与结果分析 .416.1 数据集描述与实验设置 .416.2 实验对比分析 .436.2.1 缓存击中率对比分析 .446.2.2 算法稳定性对比分析 .466.2.3 参数对 OCUC 性能影响分析 .47硕士学位论文VI6.2.4 运行效率对比分析 .506.3 本章小结 .517 总结与展望 .527.1 全文总结 .527.2 研究展望 .53参考文献 .54致谢 .59附录 攻读硕士期间发表的学术论文 .60基于用户中心点访问上下文的边缘缓存应用研究VII
