在互联网高速发展的今天，海量的信息通过互联网进行发布、传播，政治、经济、娱乐、教育、文化、学术论文等各行各业的信息，以文字、图像、声音、视频、结构化文档等多种形式进行发布，同时也在系统中进行数字化存储。信息量的不断增加提升了信息检索的难度，也对信息的有序管理提出更高的挑战。这就需要更加有效的方法和工具以高效地、自动地完成海量信息的组织、理解、检索，提高信息的利用效率和精准度，减少信息利用难度。 多标签分类是对文本、图片、视频、结构化文档等信息资源进行组织和利用的有效方法。基于多标签的自动分类技术能够提高信息的处理效率，节约人工处理成本，改善用户使用体验，近年来已得到了广泛的关注，已成为信息检索与数据挖掘研究领域的热点方向。 以LDA为代表的主题模型方法是对文本信息进行自动组织和归纳的有效方法，能够揭示文档的潜在语义，能够分析海量信息中蕴藏的主题，是重要的文本自动化处理技术，已在多标签文本分类、推荐系统等多个领域中得到广泛的应用。本文主要研究基于主题模型的有监督多标签文本分类方法和结合主题模型的个性化推荐方法。 1、经典的多标签文本分类L-LDA和Dependency-LDA模型，忽略了词项的类频率知识，即：训练数据中的一个词项被分配的标签数量。也就是说，这些模型主要关注标签内的词项权重，而忽略了标签间的词项权重。为了解决上述问题，我们试图在监督主题模型中添加一个词项权重分配的步骤，提出一种利用类频率知识对词项进行加权的方法，称作CF-weight方法：对于较低／较高类频率的词项赋予较大／较小的权重。在本文中，我们使用CF-weight方法来改进L-LDA和Dependency-LDA模型，提出WL-LDA和WD-LDA模型：在模型中，对每个词项使用其相应的CF-weight进行加权。实验结果表明，基于CF-weight的模型与现有的监督主题模型相比具有较好的分类效果。 2、个性化标签推荐的基本目的是在用户为资源进行标注时，为其提供一组候选标签，候选标签不仅与资源的内容相关，也和用户的兴趣偏好相关。本文提出一种以主题模型为基础的个性化标签推荐模型SIM-LDA-TAG：利用标签、用户、资源的关系挖掘潜在的用户兴趣主题和资源内容主题，对用户和资源进行个性化标签匹配。实验表明，将SIM-LDA-TAG的模型应用于社交分享网站进行个性化标签推荐，具有比现有主流方法更好的推荐效果。 3、个性化资源推荐的基本目的是为用户推荐一组候选资源，候选资源和用户的兴趣偏好相关。本文提出一种结合主题模型和协同过滤方法的资源推荐模型SIM-LDA：首先，利用用户对资源标注的标签，使用协同过滤进行资源推荐；其次，对用户、资源进行主题建模，通过主题相似性对用户和资源进行匹配以实现资源推荐；最后，通过权重调节混合两种推荐结果。实验表明，将SIM-LDA模型应用于社交分享网站进行个性化资源推荐，具有比现有主流方法更好的推荐效果。 关键词：主题模型；LDA；多标签分类；标签推荐；推荐系统

 Today, with the rapid development of the Internet, massive amounts of information about disseminated, political, economic, entertainment, educational, cultural, academic papers and other information are published through the Internet in many forms, such as text, image, sound, video, structured documents and so on. The increasing amount of information increases the difficulty of information retrieval and poses a higher challenge to the orderly management of information. This requires more effective methods and tools to organize, understand and retrieve massive information efficiently and automatically, to improve the efficiency and accuracy of information utilization, and to reduce the difficulty of information utilization. Multi-label classification is an effective way to organize and utilize information resources such as text, picture, video, structured document and so on. Multi-label based automatic classification technology can improve the efficiency of information processing, save the cost of manual processing and improve user experience. In recent years, it has been widely concerned, and has become a hot research direction in the field of information retrieval and data mining. The topic model method represented by LDAis an effective method for automatic analysis of text information. It can reveal the latent semantics of documents and analyze the topics contained in massive information. It is an important text automatic processing technology and has been widely used in many fields such as multi-label text classification, recommender system, etc. This paper mainly studies the supervised multi-label text classification method based on the topic model and the personalized recommendation method combined with the topic model. 1.The L-LDA and Dependency-LDA models are classic multi-label text categorization methods, but they all ignore the class frequency knowledge of the terms, that is, the number of labels assigned to a term in the training data. That is to say, these models mainly focus on the weights of the terms in the tags, while ignoring the weights of the terms between the tags. In order to solve the above problems, we try to add a term weight assignment step in the supervised topic model, and propose a method of weighting terms using class frequency knowledge, called CF-weight method: for terms with lower/higher classe frequency,give greater/smaller weights. In this paper, we use the CF-weight method to improve the L-LDA and Dependency-LDA models, and propose the WL-LDA and WD-LDA models: in these models, each term is weighted using its corresponding CF-weight. The experimental results show that the CF-weight based model has better classification effect than the existing supervised topic model. 2.Thebasic purpose of personalized tag recommendation is to provide a set of candidate tags when the user labels the resources. The candidate tags are not only related to the content of the resource, but also related to the user's interest preference. This paper proposes a personalized tag recommendation model SIM-LDA-TAG based on the topic model: exploiting the relationship between tags, users and resources to mine latent user interest topics and resource content themes, and personalize tag matching for users and resources. Experiments show that the SIM-LDA-TAG model has better recommendation effect than the existing mainstream methods when apply them to do personalized label recommendation for social sharing websites. 3.The basic purpose of personalized resource recommendation is to recommend a set of candidate resources for the user, and the candidate resources are related to the user's interest preferences. This paper proposes a resource recommendation model SIM-LDA, which combines topic model and collaborative filtering method. Firstly, collaborative filtering method is used to recommend resources by leveraging the labels assigned to resources by users. Secondly, the topic model is modeled for users and resources, and the users and the resources are matched by topic similarity to achieve resource recommendation. Finally, the two recommendation results are mixed by weight adjustment. Experiments show that the SIM-LDA model has better recommendation effect than the existing mainstream methods when apply them to do personalized resource recommendation for social sharing websites. Keywords: Topic model; LDA; Multi-label classification; Tag recommendation; Recommender Systems