社交网络中基于用户特征的专家推荐研究

发布时间：2018-08-23 09:10

【摘要】：网络社交平台经过数年蓬勃发展带动了大量用户参与,网民通过该平台与各个层面的人联系在一起,在这一过程中由微博发布、转发以及评论产生互动而形成了新的巨大信息流。这给信息获取带来便利的同时也不可避免地把信息过载这一难题推向前台,所以通过信息过滤这个手段进行个性化推荐具有重大价值。在拥有海量数据和用户的社交网络里进行信息过滤难度巨大,其中一个重要因素就在于社交网络信息发布门槛低,容易造成信息质量鱼龙混杂的现象。与此同时现有技术在分辨各个领域信息的质量高低、真假等属性的性能上还有改进的空间,因此我们可以在改进算法的同时借助特定领域的专家,依靠其专业知识和技能,帮助用户进行信息的筛选。为了达成这一目标本研究主要从用户的特征入手,依据其兴趣推荐个性化的专家。社交网络用户往往都有一些用于描述自己特征的标签,通过该标签可以快速且准确地识别用户兴趣。然而在社交网络里由于使用门槛、隐私保护等因素的限制,用户的个人标签往往不够普及,因此该特征信息稀疏,从而造成推荐困难。为解决这一问题,根据同质性,即相似的用户会喜欢相似的内容,借用用户亲密好友的特征进行标签预测。本文首先使用基于SimRank的改进算法ASCOS对用户社交关系相似度进行计算,然后进行两两比较找出用户的亲密好友,再根据其好友标签进行预测。随后在专家识别方面提出了依据PageRank为原型的FRank算法,改进了原始算法在小社交圈内计算不准的缺陷。实证表明,使用ASCOS在用户标签预测中的准确率和召回率上得到提升。在专家预测上使用nDGC作为性能评价标准,并发现与基线方法PageRank相比FRank的性能也有所提升。最终依据上述两个方面的成果,即标签预测和专家识别,实现了向用户推荐个性化的专家。本文共分6章:第1章,介绍当前社交网络个性化推荐的研究现状和成果,描述文章结构。第2章,说明本文研究所需的理论技术,包括数据采集的方法、用户标签、用户特征模型及常见推荐系统。第3章,根据新浪微博用户的宏观特点选择用户标签作为推荐模型的基础,在此基础之上使用ASCOS算法计算用户相似度并对用户标签进行扩展。第4章,通过预测标签和社交关系网络,使用FRank算法对该网络中的专家用户进行识别,并完成推荐系统设计。第5章,对新浪微博的数据进行分析,得到标签预测和专家识别的结果,通过实证验证本研究的有效性。第6章,总结分析结果,找出本研究所存在改进的空间,为下一步的探寻做好铺垫。
[Abstract]:After several years of vigorous development, the social networking platform has brought a large number of users to participate, through which Internet users connect with people at all levels, and in the process, they are released by Weibo. The interaction between retweets and comments creates a huge new flow of information. This brings convenience to information acquisition, but also inevitably brings the problem of information overload to the foreground, so it is of great value to carry out personalized recommendation through information filtering. It is very difficult to filter information in social networks with huge data and users. One of the important factors is the low threshold of information release on social networks, which can easily lead to mixed information quality. At the same time, there is still room for improvement in the existing technology in terms of distinguishing the quality of information in various fields and the performance of attributes such as truth and falsehood, so we can improve our algorithms while relying on the expertise and skills of experts in specific fields. Help users filter information. In order to achieve this goal, this study mainly starts with the characteristics of users and recommends personalized experts according to their interests. Social network users often have tags to describe their own characteristics, which can quickly and accurately identify user interests. However, due to the restrictions of threshold and privacy protection in social networks, the user's personal tags are often not popular enough, so the feature information is sparse, resulting in the difficulty of recommendation. In order to solve this problem, according to homogeneity, that is, similar users will like similar content, and use the characteristics of close friends to predict the labels. In this paper, we first use the improved algorithm ASCOS based on SimRank to calculate the similarity of users' social relations, and then compare and find out the close friends of users, and then predict them according to their friend tags. Then the FRank algorithm based on PageRank is put forward in the aspect of expert recognition, which improves the defects of the original algorithm in the small social circle. The empirical results show that the accuracy and recall rate of user label prediction are improved by using ASCOS. In the expert prediction, nDGC is used as the performance evaluation standard, and the performance of FRank is also improved compared with the baseline method PageRank. Finally, according to the results of above two aspects, label prediction and expert recognition, personalized experts are recommended to users. This paper is divided into six chapters: chapter 1 introduces the current research status and achievements of personalized recommendation of social networks, describes the structure of the article. In chapter 2, the theory and technology needed in this paper are described, including data acquisition method, user label, user feature model and common recommendation system. In chapter 3, according to the macro characteristics of Sina Weibo users, the user tags are selected as the basis of the recommendation model, and the ASCOS algorithm is used to calculate the user similarity and extend the user tags. In chapter 4, the expert users in the network are identified by using the FRank algorithm and the recommendation system is designed by using the prediction label and the social relationship network. In chapter 5, the data of Sina Weibo are analyzed, the results of label prediction and expert identification are obtained, and the validity of this study is verified by empirical analysis. Chapter 6, summing up the analysis results, find out the room for improvement in this research, and pave the way for the next step.
【学位授予单位】：华中师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：G252;G206

【参考文献】