基于社会化媒体的观点和行为挖掘研究

发布时间：2018-05-27 14:44

本文选题：观点分析 + 行为分析　；参考：《北京邮电大学》2014年博士论文

【摘要】：随着Web2.0技术的发展,互联网在线服务已经逐渐变成人们日常生活中不可或缺的组成部分。网络使用过程中所产生的文字、图像、音频、视频以及服务使用记录等共同构成互联网UGC(用户生成内容)的海洋。一方面,新技术的应用和流行能够为人们日常生活带来诸多便利,因而激发用户需求的日益膨胀。另一方面,新技术的诞生、成熟以及应用阶段也面临着各类挑战和困难,从而产生诸多亟需解决的问题。本文从四个方面对社会化媒体上的用户观点和行为进行分析和研究。第一,提出一种基于反义距离概念的反义疑问观点的特征构建方法以及观点文本建模方法。由于互联网文本具有海量,不规范,短文本等特性,本文提出用反义距离RDT概念筛选语言模式来构建特征库,并利用组特征GF计算文本和特征库之间的相似程度。验证实验结果表明该方法在多种机器学习分类算法下能够达到较高的识别准确率,同时由于特征向量被降维,能够减少建模的时间开销。文中亦对平滑因子以及反义特征库筛选阈值参数对识别准确率的影响进行分析和讨论。第二,提出一种基于结构性上下文的主观表述库扩展方法。由于基于词典的扩展方法新词发现能力受限,扩展规模较小,依赖分词和词性工具等不足；另外,基于语料库的利用连接词语言规则和基于共现规则发现主观表述的方法存在对备选主观表述覆盖度低的不足。因此,本文选用结构上下文来指导主观表述库的扩展,它利用信息量和点互信息PMI概念来度量结构上下文对于种子词汇出现事件的预测能力,以及上下文和备选表述的使用环境相似性,并计算和推断备选表述的主观性指标值。验证实验结果表明该方法能有效地构建精准,新颖主观表述库,并能充分补充其它方法生成的表述库。第三,提出一种计算知识分享社区中用户知识贡献能力的方法。传统知识分享网站由于社交特性的加入,知识从原本存储于服务器上的静态内容转变分散在各个潜在用户处的动态知识。服务核心任务由查询-内容相似性识别转变为潜在用户知识提供能力度量上。本文对用户活跃度,内容质量以及社会影响力几个角度的能力指标进行分析和量化,通过改进的社会网路链接分析方法来得到综合指标,生成系统推荐用户列表。此外,文中还对全网迭代值的计算收敛性进行分析和研究。验证实验结果表明知识贡献能力模型能够有效为用户的各方面能力进行综合建模,防止单一指标出现的筛选偏差。文中对衰减因子对用户能力筛选倾向性影响以及计算收敛性的影响进行分析和讨论。此外,文中还提供不同能力指标对于综合能力贡献权重的分配方案。第四,提出一种计算互联网微博用户媒体信息能力的方法。微博上海量用户帐号给新注册用户筛选待订阅的信息源帐号带来障碍,本文提出媒体源能力模型来衡量微博账号提供新闻内容的能力。该方法量化用户的个人活跃度、用户所提供内容的可信程度、用户所提供的内容量、用户稳定输出内容的能力。通过集中意见模型加权波达排序方法将不同能力空间的排序结果映射到统一的度量指标上,得到用户媒体能力的最终排序结果。验证实验结果表明媒体能力模型能够有效刻画用户提供媒体资讯内容的能力,遴选出综合能力表现突出的账号。
[Abstract]:With the development of Web2.0 technology, Internet online service has gradually become an integral part of people's daily life. The words, images, audio, video and service records produced in the process of network use constitute the ocean of the Internet UGC (user generated internal capacity). On the one hand, the application and popularity of new technology On the other hand, the birth, maturity and application of the new technology are facing various challenges and difficulties, and there are many problems to be solved. This paper analyzes and studies the views and behavior of the user in the social media from four aspects. Study.
First, a feature building method of antisense query based on antisense distance concept and a method of view text modeling are proposed. Because the Internet text has the characteristics of mass, nonstandard, short text and so on, this paper proposes an antisense distance RDT concept screening language model to construct the feature library, and uses the group feature GF to compute the text and feature library. The experimental results show that the method can achieve higher recognition accuracy in a variety of machine learning classification algorithms. At the same time, the time overhead of modeling can be reduced because of the feature vector being reduced. The effect of filtering threshold parameters on the recognition accuracy is analyzed and the effect of the filtering threshold parameters on the smoothing factor and the antisense feature library is also analyzed. Discuss.
Second, a subjective expression library extension method based on structural context is proposed. Due to the limited ability of the new word discovery based on the dictionary extension method, the expansion size is smaller, the word segmentation and the word based tools are not enough. In addition, the method of using the corpus based language rules and the co-occurrence rules to find the subjective expression is opposite. Therefore, this paper uses the structure to guide the expansion of the subjective expression library. It uses the information quantity and the point mutual information PMI concept to measure the prediction ability of the structure context to the occurrence of the seed vocabulary events, and the context and the alternative table of the use of the environmental similarity, and calculate and deduce the alternatives. The experimental results show that the method can effectively construct the precision, the novel subjective expression library, and can fully supplement the expression library generated by other methods.
Third, a method of computing knowledge sharing the knowledge contribution ability of the user in the community. The traditional knowledge sharing website, due to the social characteristics, changes the dynamic knowledge from the static content originally stored on the server. The service core task is transformed from the query to the content similarity recognition to the potential. This paper analyzes and quantifies the ability indexes of user activity, content quality and social influence, and obtains the comprehensive index by improving the social network link analysis method, and generates the list of recommended users. In addition, the convergence of the iterative value of the whole network is also carried out in this paper. Analysis and research. The experimental results show that the knowledge contribution capability model can effectively build a comprehensive model for all aspects of the user's ability, and prevent the screening deviation of the single index. In this paper, the influence of attenuation factor on the user ability screening tendency and the influence of calculation convergence are analyzed and discussed. In addition, the difference is also provided in the paper. The allocation scheme of ability index for comprehensive ability contribution weight.
Fourth, a method to calculate the information ability of Internet micro-blog user media is proposed. The user account number of micro-blog Shanghai is an obstacle to the new registered user to select the information source account to be subscribed to. This paper proposes a media source capability model to measure the ability of the micro-blog account to provide news content. This method quantifies the user's personal activity, and the user proposes The credibility of the content, the internal capacity provided by the user, the ability of the user to stabilize the content of the content. By mapping the sorting results of the different capacity spaces to the unified metric by the weighted Boda sorting method of the centralized opinion model, the final sorting result of the user media ability is obtained. The experimental results show that the media capability model can be used. It effectively depicts the ability of users to provide media information content and selects outstanding accounts with comprehensive capabilities.
【学位授予单位】：北京邮电大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TP393.09;TP391.1

【共引文献】