The Research Of Recommendation Model Of Electronic Commerce Fused By BP Network

The development of computer science and Internet,make profound changes in allaspects of people’s daily lives,and the network information available which showsexplosive growth at the same time,is far beyond the amount that people canaccept,leading to information overload.Due to a wide variety and a large number ofitems in E-commerce Website,users tend to be inundated with large amounts ofinformation,so that they can not fully explore their favorite items.This leads to thephenomenon that80%of sales come from20%of the merchandise.Personalizedrecommendation is an effective way to solve this problem.This article firstly gives a brief description about the architecture and the workingprinciple of personalized recommendation system,and analysis the limitations of thetraditional content-based recommendation algorithm and neighborhoodrecommendation algorithm based on memory,when faced with a huge amount of usersand a wide variety of goods in e-commerce application scenarios.On this basis,fromitem properties,we utilize machine learning and the optimization theory to analyze theLFM model based on singular value decomposition,and then combined with thespecific circumstances of the user’s shopping in e-commerce website,we put differentuser behaviors generated from shopping process corresponding to different implicitfeedback information,and map them to different scoring proportion.We use thismethod to improve the existing LFM models which consider only a single implicitfeedback information, so that it can reduce the rating prediction error.Then, from the perspective of user interest groups, We applied LDA model basedon probability distribution in text mining area to rating prediction problems,whichachieved a good prediction effect.On the basis of the work above,in view of theperformance bottlenecks of the single recommendation model,this paper use artificialneural network to integrate the LDA and improved LFM model,to explore theinfluence of multi-model fusion for performance improvement of therecommendation.Experiments show that in the rating prediction problem,the fusionmodel can achieve lower RMSE value than the single model.subsequently,this articletransform rating prediction problem into actual Top N recommendation problem.Theexperiment proved that the RMSE decrease slightly, but it can bring a significantimprovement in the accuracy of recommendation.Finally, we summarize the research work in this paper,and make a prospect forthe study of other aspects of the personalized recommendation.