Research On Behavior Recommendation Model Based On Neural Network

With the rapid development of Internet technology,Internet technology applications such as Amazon,e Bay,Tmall,Meituan and Elm are becoming an important part of people's daily lives.While these applications provide convenient services for people,they also produce a large amount of user-item,user-user and item-item interaction data,which often hides a large number of user preference information and user behavior semantics.By mining this massive amount of information,we can quantify a user's preferences,the properties of an item,and the impact of information dissemination between users and users.Therefore,we can put semantic information into personalized recommendation,public opinion monitoring,user portraits and accurate advertising matching and other fields.In the traditional recommended field,the existing recommendation model which based on users' score cannot solve the cold start problem caused by the sparse data.There are abundant interactive data between users and items,but traditional recommendation system ignores the interactive data of other user-items,such as comment information,the order of browsing or purchasing items and the interactive information of user's social relationship.Compared with the traditional recommendation field,POI recommendation faces a huge problem of processing different types of data.Therefore,the data sparsity of POI recommendation is higher than that of traditional recommendation,and the cold start problem is more serious.Therefore,the Next-POI recommendation system based on the comprehensive utilization of LBSN data has become a hot research topic in academia and industry in recent years.In view of the above two problems in the field of recommendation system,this paper studies from the following two aspects:(1)Attention Neural Network for User Behavior Modeling.We proposed ATUBM(Attention Neural Network for User Behavior Modeling)method to capture the preferences in user reviews and predict the next product that users might want to buy based on their purchase behavior.The main idea of AT-UBM is to use the neural network to filter all users' reviews,sample out the representative review information,input the review information into a attention neural network to assign different weights,and finally select the representative word vectors to represent the features of users and the features of items respectively.AT-UBM will concat the review features of items according to the order of user access,and finally use CNN neural network to extract user features and make predictions.The experimental results of the real data set show that the HR of ATUBM has an increase of 6% to 8% compared with the AT-Rank.(2)Social-geographic Behavior Alliance model.The LBSN graph is a heterogeneous graph,which contains many data with different structures.The existing LBSN-based POI recommendation system usually divides LBSN into multiple isomorphic for learning,which will result in the loss of a lot of heterogeneous information.This paper proposes a heterogeneous graph embedding learning model SGBA(Socialgeographic Behavior Alliance model)to directly learn the POI embedding representation on LBSN graph.SGBA first obtains each POI associated POI sequence by heterogeneous random walk algorithm,and then generates POI embedding by Word2 vec.SGBA divided different access blocks according to the time and geographical distance of users accessing the POI and used LSTM to learn the preference of user access sequence from different session scales.Experimental results show that compared with other algorithms,SGBA has better performance in AUC,Recall and ACC.Among them,AUC performance improved by 3.4% compared to ST-RNN on Gowalla,and 3.2% in Brightkite compared to ST-RNN.