Research Of Rank Based Author-topic Model

With the prevailing of Web2.0applications, more and more web users are activelypublishing text information online, which exceeds the capabilities of both documents andindividuals. These users also often form social networks in various ways, leading tosimultaneous growth of both text information and network structures such as socialnetworks. Taking academic papers as an example, as researchers are regularly publishingpapers, we not only obtain text information, but also naturally have availableco-authorship networks of authors. One can easily imagine many other examples of textaccompanied by network structures such as webpages accompanied by links and literatureaccompanied by citations.Topic modeling becomes a widely used tool for managing large volume of documents,such as document categorization. However, previous topic modeling methods do notdistinguish the importance of documents on diferent topics while performing topicmodeling. By incorporating the link based ranking, we establish a novel topic modelcalled rank based topic model. Topical pagerank is used to calculate the topical rankingswhich indicate the importance of documents on topics. The rank based topic model can beapplied to retrieve great papers of a particular area and classify papers into differentcategories according to the topics.Author-topic is widely used that to extract both the topics and authors’ interests fromdocument collections. State-of-the-art author-topic models consider authors are equallyimportant to all the topics. However, intuitively, authors have different degrees ofimportance on different topics in real life. Ignoring such feature may inherently hurt theperformance of author-topic models. In this paper, we propose a rank based author-topicmodel that incorporates the link based rank into author-topic model. We exploit topicalpagerank to calculate the topic specific rank of authors which indicates the importance ofauthors on specific topic. Rank based author-topic model combines topic-sensitive rankand author-topic model, and leverages the power of rank to improve author-topic model.Rank based author-topic model can be applied to retrieve famous researchers of aparticular research area and suggest reviewers for papers.Based on the proposed models, experiments are implemented on a subset of theArnetMiner data. Experimental results show that rank based author-topic modeloutperforms author-topic model and iTopic in terms of perplexity. Furthermore, we qualitatively and quantitatively demonstrate that topics detected by rank basedauthor-topic model are more interpretable and semantically coherent than those detectedby traditional author-topic models.