基于用户兴趣的微博个性化信息推荐研究

发布时间：2019-03-27 09:17

【摘要】：随着移动智能和互联网的飞速发展，人们从信息匮乏的年代过度到了信息过载的时代。微博作为一种新型的社会化自媒体平台，近年来用户数量呈指数增长，每天生成大量的UGC（User Generating Content）。如何挖掘用户的个人兴趣建立用户兴趣模型，并将用户感兴趣的信息从海量信息中找出推荐给用户显得尤为重要。本文以微博用户的兴趣建模和微博个性化信息推荐为研究内容。主要包括：（1）传统的向量空间模型和TF-IDF方法没有考虑语义信息且存在用户特征高维稀疏的问题，而常用的基于文档级别词共现的潜在狄利克雷分配模型（Latent Dirichletallocation,LDA）并不适用于微博这种短文本的主题挖掘和用户兴趣建模。鉴于此，本文引入适用于短文本的主题模型BTM（Biterm Topic Model）挖掘用户的个人兴趣，结合用户兴趣的多变性，提出基于时间窗口的用户动态兴趣模型。（2）在用户兴趣模型的基础上，，针对微博中用户收听列表信息过载的问题，提出综合考虑微博本身质量、用户个人兴趣和社交兴趣这三个主要特征的推荐模型，并在模型中引入协同过滤的思想。针对微博中用户主动获取的其他信息（非用户收听列表的信息），提出一种基于主题的信息推荐思想，并以美食主题为例，设计了整个应用。（3）通过Big Data平台获取实验数据，通过实验验证了BTM建立的用户兴趣模型在推荐性能上要优于LDA及TF-IDF模型且考虑用户兴趣的多变性能进一步优化推荐效果；在三个主要影响因素中，结合了协同过滤思想的用户个人兴趣特征推荐性能最优，用户社交兴趣特征次之，微博本身质量特征最差；本文提出的推荐模型从用户兴趣建模出发，针对不同的场景结合不同的特征构建推荐模型，任何UGC平台的信息推荐问题都能够在本文的研究基础上进行扩展利用。
[Abstract]:With the rapid development of mobile intelligence and Internet, people from the era of lack of information to the era of information overload. Weibo, as a new type of social self-media platform, has seen an exponential increase in the number of users in recent years, generating a large number of UGC (User Generating Content). Every day. How to mine the user's personal interest to establish the user interest model, and find out the information that the user is interested in from the massive information to recommend to the user is very important. This article takes Weibo user's interest modeling and Weibo personalized information recommendation as the research content. The main contents are as follows: (1) the traditional vector space model and TF-IDF method do not consider semantic information and have the problem of high-dimensional sparse user characteristics. However, the commonly used latent Dirichlet allocation model based on document-level co-occurrence of words (Latent Dirichletallocation, LDA) is not suitable for topic mining and user interest modeling of short text such as Weibo. In view of this, this paper introduces a topic model, BTM (Biterm Topic Model), which is suitable for short text, to mine users' personal interests. Combined with the variability of user's interests, a dynamic user interest model based on time window is proposed in this paper. (2) on the basis of user interest model, aiming at the problem of information overload in Weibo's listening list, a recommendation model considering Weibo's own quality, user's personal interest and social interest is put forward. The idea of collaborative filtering is introduced into the model. Aiming at the other information (non-user listens list information) obtained by users in Weibo, this paper puts forward a subject-based information recommendation idea, and designs the whole application with the gourmet theme as an example. (3) get the experimental data through the Big Data platform, and verify that the user interest model established by BTM is better than the LDA and TF-IDF model in recommendation performance, and further optimizes the recommendation effect considering the changeable performance of user interest. Among the three main influencing factors, the recommended performance of user's personal interest characteristics combined with collaborative filtering is the best, that of user's social interest is the second, and Weibo's own quality is the worst. The recommendation model proposed in this paper starts from the user interest modeling and constructs the recommendation model according to different scenarios combined with different features. Any information recommendation problem on UGC platform can be extended and utilized on the basis of the research in this paper.
【学位授予单位】：华南理工大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP391.3;TP393.092

【参考文献】