Mining User Dynamic Interest Topic For Social Tagging

In the social tagging system,the user can freely define tags to the resources of interest based on their own understanding.As the keywords selected by users,tags reflect the user's views and interests on information resources.The discovery and acquisition of user interest is the core and foundation of personalized recommendation service,while the existing researches mainly focus on the relationship between users and tags in the way of the quantity and structure.Moreover,the actual situation is that the tag is only a single vocabulary that reflects the interest of the user,the user interest is often the theme of the collection formed by multiple tags and the characteristics of the tag dynamic change with the user's tagging time will also affect the portrayal of user interest.Therefore,combined with the Natural Science Fund Project “Research on social-driven context-aware personalized information service in ubiquitous computing environment”(project NO: 71471165),this paper presents the mining of users' interest topics in social tagging system,which includes expatiating on the construction of user interest topic model systematically,putting forward dynamic tag correlation indices to mind real user interest topic and carrying out experimental study.The main contributions of this thesis are listed as follows:(1)The topic model is constructed to obtain the tag topic.Based on the "user-tag" binary relation,the user tagging information is regarded as a corpus,and the tag is regarded as the word in the corpus,so as to construct the LDA model to explore the potential tag topic.In this way,the original disorganized snd dynamic tags are divided into clusters with different topics.It can not only overcome the uncertainty and inconsistency of the tags,but also improve the accuracy and efficiency of the tag information effectively.(2)Dynamic association indices are proposed to obtain user tag interest.Establish an association space model to describe the relationship between the user and his tag set firstly,then by analyzing the tagging process with time series and considering the time characteristics of tags,the paper defines intensity index reflecting the quantitative characteristics and stability index reflecting the time change.Lastly based on the indices above,the correspondence between user tag set and user interest topic is realized effectively.(3)Mining dynamic user interest.Combining the tag topic and the user's dynamic tag interest,the similarity between them is calculated,and the user's dynamic interest topic can be obtained.(4)An experimental study is carried out with the real user annotation data from Last.fm social annotation platform to mine user interest topic.We use the coverage and accuracy criterions of Acc and Rec to validate the validity of the user interest topic model based on the dynamic correlation indices proposed in this paper.The results show that the proposed user interest topic based on dynamic correlation indices has better performance in terms of validity and is superior to the mining measures based on TF method and TF-IDF method in both coverage and prediction accuracy.This research result has great pratical application value for personalized recommendation based on social tag.