Cross Domain Product Recommendation Based On Review Text

With the rapid development of information technology and the population of PCs and mobile devices,it's more and more convenient for people to shopping through e-commerce platform,which also makes the e-commerce platform continue to expand,bringing consumer a wide range of products,but also the problem of information overload.In the past ten years,recommender system as an important way to solve information overload,has achieved certain success in the field of e-commerce.While most of them also inevitably suffer from the data sparseness and user cold-start problem.The improvement of material living standards has made people's demand for shopping more diversified,refined and personalized,which has led to the development of e-commerce platform in the direction of refined operation.The main e-commerce platform continues to improve the commodity field,providing consumers with more accurate product services,and constantly enriching the way users feedback,such as review text,not only just the simple rating feedback.This makes the e-commerce platform draw their user profile more and more precisely,which provides them a possible approach to solve the data sparsity and user cold start problems in the traditional recommendation.Based on the above two points,this paper made some researches on data sparseness and user cold start problem from the perspective of cross-domain recommendation and review text mining.Two cross-domain user preference model were proposed by review text mining,which were used to develop two product recommendation algorithms for difference recommendation scenarios.First,this paper developed CCoNN leveraging the user preference information shared among domains to solve the data sparseness problem.CCoNN is based on CNN,extracting user preference vector shared among different domains and domain-specified product feature vector from review text.Such vector was leverage by Factorization machine to improve the precision of rating prediction.Second,this paper focus on solving user cold-start problem by migrating their preference information from other product domain,which led to DSNRec.DSNRec is a encoder-decoder structure,which sperate the domain-common and domain-specified user preference information from review text to migrate the shared user preference among difference domain.It also generates product feature vector from review text.Finally,the domain-common user preference information and product feature vector are combined by Matrix Factorization to generate rating predictionFinally,experiments are designed on the Amazon dataset to verify the effectiveness of CCoNN and DSNRec.