The Study Of Recommendation System

The rapid development of Internet and information technology facilitates the production and dissemination of information,brings great trouble to the consumption of information as well.The consequent "information overload" is causing information producers to worry about how to make their products stand out,while also worrying information consumers about how to get the information they want.Search engine,represented by Google and Baidu can alleviate the difficulties brought by" information overload" to a certain extent.But search engine is also not possible when it comes to the questions like personalized demand,automatic matching and ambiguity.Recommendation system was born under such circumstances.Recommendation system has achieved great achievements so far.The classic recommendation algorithms like CF,MF and LFM have achieved amazing results on amazon,Netflix,Taobao,doudou and other major websites.However,recommendation system still suffers from urgent problems such as cold start,data sparsity,and system precision and diversity.A good recommendation system needs to mitigate or avoid these problems as possible.In this thesis,the problems of subjective and complex of UGC and the diversity of traditional recommendation algorithms are introduced,and put forward some novel algorithms to avoid these problems.The main research work is as follows:The research background and significance of recommendation system,the focus in the field of recommendation system and some common classical methods is briefly introduced.The user participation problem and label validity problem of UGC is analyzed.For these problems,this thesis presents a novel method called IAUM to finish item modeling and user modeling based on the high efficient attribute label.Based on IAUM,the recommendation system structure was designed.IAUM mainly adopts One-Hot encoding and LWSM to complete the modeling.Experiments on MovieLens 1M show that the IAUM has higher precision than MF.The low diversity of traditional recommendation methods,the need of improving the users' experience and some existing improvement methods is introduced.Two matrix factorization algorithms,RDMF and DMF are proposed.This two methods optimize precision and diversity in the decomposing process by independently drawing the diversity influencing factors.The experimental results show that the recommendation system based on RDMF and DMF is much better than the recommended system based on MF,and the recommendation results are more consistent with the diversity preferences of individual users.