Research On Recommendation Algorithm And Recommendation System For E-commerce

With the rapid development of the Internet?the popularity of smart phones and Improvement of communication infrastructure,e-commerce has become an indispensable part of people's lives.From the rise of various e-commerce website and the annual large-scale promotional activities,we can see the big potential in e-commerce.Not only the number of users is increasing but a variety of different quality of the goods are also filled with the entire e-commerce field.This.-leads to that users have to spend a lot of time to choose their interested products from huge amount of goods.How toquickly find out the products that people are interested in from thousands of products has become an important topic in academia and industry,and the personalizedrecommendation system in e-commerce is an important way to solve this problem.According to the previous research on the personalized recommendation system,this article takes the e-commerce platform as the research object,and tr es to put forward a new algorithm to enhance.the platform user experience as well as the commercial value Firstly,it summarizes the origin and development of recommendation system by using the literature research method,and describes the algorithms which are commonly used in the recommendation system.Based on the traditional cooperative filtering algorithm,it puts forward a new algorithm based on network sparseness and the fusion.of time attribute and label information,and the algorithm is evaluated by offline experiment and online experiment.The experiment shows that the algorithm has a 5.2%improvement compared with the original algorithm.At the same time,the Wide&Deep model was applied to the field of e-commerce,and the Wide&Deep model based on TensorFlow is designed and implemented.In the online experiment,the click rate is increased by 8.3%.The main research contributions of the article are followings:(1)The general framework of the recommendation system in the industry is proposed.The recommendation system is divided into four modules:Match stage,Rank stage,ReRank stage,and display stage.Among them,the core of the two modules are the Match stage and Rank stage,and the recommended algorithms are usually applied in these two modules.The proposed framework allows us to clearly understand how the various parts of the recommended system work,in order to deepen our understanding of the recommended system.(2)This paper gives a complete overview of personalized recommendation technology,and introduces the algorithms of Match and Rank.The Match algorithms section introduces the traditional collaborative filtering algorithms,and the graph-based link prediction algorithms.In Rank algorithms section,it introduces several commonly used linear model recommendation system(logistic regression model),nonlinear model(GBDT)and deep learning model(Deep Neural Network,Recurrent Neural Network).At the same time,it points out the existing problems and the advantages of the proposed algorithm.(3)An improved Match algorithm is proposed:By adjusting the network sparseness,considering the time attribute,introducing the tag information to solve the problem of low confidence,using the hierarchical i2i and the horizontal and vertical normalization methods,the algorithm can be applied to the large-scale recommendation system,while ensuring the diversity and accuracy of the recommended results.(4)The Wide&Deep model is applied to the Rank stage in the e-commerce recommendation system.The work of the feature engineering is introduced in detail,and the method based on the multi-dimensional to get features is proposed.This paper summarizes the architecture of the recommendation system,and discusses it from the aspects of sample generation,model training and online service.This model combines the advantages of logistic regression model and deep neural network model to achieve both memorization and generalization.