Research On E-commerce Platform Recommendation Method Based On User Preference Behavior

With the development of society,people's pursuit of quality of life has increased rapidly.As an online virtual expansion of physical services,network services have brought tremendous convenience to people's lives.Herein,e-commerce,as an important component of network services,is increasingly sought after by the people.However,everything has two sides.While enjoying network services to improve living standards,they are also facing the issue of selectivity brought about by massive amounts of information.In this context,the recommendation system,as an important tool to assist netizens in selecting high-quality content,recommends items that are likely to generate interest to users,and has always been a hot field of research.This thesis summarizes the current research status of recommender systems in recent years,mainly from the perspective of user group interest behavior.Through the analysis of the user's historical behavior data,on the one hand,this thesis studies recommendation algorithms based on user interest group discovery and data enhancement,and on the other hand studies recommendation based on representation learning and matrix factorization.The main research contents are as follows:1.User interest group identification and generation algorithm against network recommendation.This thesis proposed a recommendation model based on data compensation and user dynamic interest group.First,by introducing the advantages of the generative adversarial network in learning data distribution and enhancing data samples,homomorphic compensation is performed on the original data,and the preference relationship between users and projects is more truly restored.Secondly,for the generalization of user interest,information entropy is introduced to measure user interest feature space.At the same time,around the problem of user interest drift,the time window identification method is used to further quantify user dynamic interest groups.Finally,in order to integrate the feature vectors of user interest groups,a third-order tensor of "user-item-interest group" is constructed,and missing values are calculated during the decomposition process to predict user ratings more reasonably.2.Recommendation algorithm based on representation learning and matrix factorization.There is a large amount of auxiliary information in the heterogeneous information network,which has a great effect on data analysis tasks.Firstly,how to extract effective information from the network,this thesis introduces the concept of meta path and designs a wandering strategy,which can better extract the node sequence used for network representation.Then,considering that the semantics of each node vector is different under different element paths,and considering the expansibility of matrix factorization,a set of functions are designed to map the expressed nodes into the matrix factorization model.Finally,this paper uses matrix decomposition to predict the user's score of the item.In this thesis,five public datasets are used to analyze the proposed method.Experimental results show that the proposed model can improve the recommendation accuracy problem caused by user's no behavior preference and data sparsity.