Recommendation Algorithms By Exploiting Rating Matrix And Review Texts

Recommendation system exploits user-generated information(User generated content,i.e.,UGC)has been researched and applied generally.The content based on ratings exists widely in numerous UGC.Therefore,recommendation algorithms by exploiting rating matrix have attracted more attention,for instance,collaborative filtering algorithm,clustering,association rules,hidden factor model.With the successful application of recommendation system,more and more users participate actively in the work;so the increasing information is applied to the recommended algorithm constantly.The reviews written by users have become one of the most influential information,which can reflect the user preferences.The researchers also take more and more concerns on these reviews,So that they can improve and optimize the text analysis and opinion mining method more effectively.However,recommendation algorithms by exploiting rating matrix and review texts are very rare,so we have to improve and put forward four type fusion rating matrix and review texts information recommendation algorithms to improve the effect of recommendation and identify the useful reviews,which based on the basis of previous studies.The main work is as follows:(1)Rich users' emotion and abundant product features have been contained in these reviews,but the extant methods couldn't combine ratings and reviews ideally.The thesis proposes two optimized methods: HFPT algorithm and DLMF algorithm,which intends to improve the HFT(item)algorithm.The thesis takes an experiment on 28 groups' data from Amazon,and regards average mean square error(MSE)as the indicator.As a result,we found that HFT(item)algorithm is better than HFPT algorithm.Because the contents of a single review are scanty,Latent Dirichlet Allocation algorithm for theme discovery applied in short text is inferior to long text.So DLMF algorithm divided review collections into user review collection and product reviews,which can reflect the user preferences and product features more concretely.Trying to merge them into the matrix factorization model to solve above problems,eventually,in the term of mean square error(MSE),we found that DLMF algorithm is superior to the HPT(item)algorithm,and the enhanced effect of data subset has been up to 3.68%.(2)User's affection can affect the commodities preferences of them and others,regarding the theme preferences as a guide are beneficial to enhancing effect.The thesis proposes two modified algorithms by taking theme preferences guide into account: PGMF algorithms and DPGMF algorithms.In the term of MSE,the experiment found that the two algorithms are be superior to HFPT ? DLMF ?HFT(item).Especially,compared with HFT(item),the enhanced effect of data subset has been up to7.31%.(3)In the previous studies,the single commodity was regarded as the indicator,and sorting all reviews belongs to the items.The thesis put forward identifying the usefulness of personalized users' reviews,which is on the basis of the above four type algorithms.And we can get the sorted reviews,which is the most relevant to user preferences and commodity characteristics.And provide the final results for different users.However,the present data set do not provide a useful marker for personal user,so it cannot give the effects of reviews' sort,but you can do research on possible data in follow-up study.