Multiplex Topic Models And Dual Matrix Factorization For Recommendation Systems

With the fast development of internet, the information structure has become various.Document contain not only content information but also relational information such as coau-thors, citations and georagphic locations, leading to the multiplex network structure. Intu-itively, different types of links may carry different kinds of semantic meanings, for example,two authors who often collaborate on some papers tend to be interested in the same top-ics, while the cited paper may be from an interdisciplinary area with a quite different topic.How to quantify and balance different types of links in document networks for a better topicmodeling performance still remains a challenging and unsolved problem.In this thesis, we propose a novel multiplex topic model (MTM) that represents thetopic influence from different types of links using a factor graph. To estimate parameters inMTM, we also develop an approximate inference algorithm, referred to as multiplex beliefpropagation (MBP), which can estimate the influence weights of multiple links automati-cally at each learning iteration. Experimental results confirm the superiority of MTM in twoapplications, document clustering and link prediction, when compared with several state-of-the-art link-based topic models.Collaborative filtering is one of the most successful technologies in the recommenda-tion system. Matrix factorization mode is one of the most widely used algorithms in col-laborative filtering system. A large number of studies have shown that matrix factorizationmodel is superior over other collaborative filtering algorithms in terms of the recommen-dation speed and accuracy. While matrix factorization model relies on the user-item ratingmatrix, if a new item is added to the system and there are no ratings to it, or a new usergives no ratings to all the items, matrix factorization cannot work at this moment. This isthe common cold-start problem in recommendation systems.In this thesis, we propose a novel recommendation technique based on dual matrix fac-torization (DMF) to solve the cold-start problem in collaborative filtering systems. DMFcombines PMF and ATM to factorize the rating matrix and word-author matrix simulta-neously. The experimental result shows that the proposed algorithm not only shows good performance on the in-matrix prediction, but also has10%recall achievement on the out-of-matrix prediction when compared with CTR.