Research And Implementation Of Recommendation Technology Based On Multi-scene Session Data

With the rapid development of information technology and Internet in recent years,the problem of information overload is becoming more and more serious.It becomes very difficult for users to find information suitable for their personalized needs from the vast resources of texts,videos,images and commodities.Personalized recommendation system has become an important means to solve this problem.Compared with search engines,recommendation system can obtain users' interests and preferences based on the research of users' historical behavior data and statistical analysis modeling,so as to guide users to discover their information needs and realize personalized recommendation.Therefore,this technology has been widely used in the Web3.0 technology based Internet platform.This thesis intends to study cross-domain recommendation methods for partial overlap between fields.Since the user sets in different domains overlap completely and the user sets in different domains do not overlap completely are two extreme cases,there is more partial overlap between the user sets in real life.Because many websites now provide access to other accounts,from this point of view,you can find the same user in different fields.In addition,experiments have proved that this small number of overlapping users have interacted with over 80% of items in each domain.It is reliable and effective to use this information as a bridge for information sharing and migration between domains.The background of this thesis is an information system based on multiscene heterogeneous data of single system and multiple scenes,and the users of each scene are almost the same.Therefore,the cross-domain recommendation method based on user overlap is reliable and effective.Based on graph neural network and probability matrix decomposition,this thesis designs and implements a collaborative graph neural network(CGNN),which uses Node2 vec as the global object feature vector extraction model,PMF as the local user and object feature vector extraction model,and uses standard Bayesian Personalized Ranking(BPR OPT)as an optimization method.Through the experiments in the open real data set,it is proved that the model recommendation results in this thesis are better than the comparison model recommendation results.The collaborative graph neural network algorithm proposed in this thesis effectively solves the problem of object vector expression and user vector expression in multi scenes,and effectively improves the quality of recommendation.At the same time,the collaborative graph neural network model is deployed in the actual application system,and good performance is achieved.The main contents of this thesis are as follows:1.A recommendation algorithm CGNN combining graph neural network and probability matrix decomposition is designed and implemented2.CGNN was applied to the multi-scene recommendation problem to solve the cross-domain recommendation problem of partial overlap between domains under multiscene,and a good effect was obtained3.Designed and implemented a multi-scene session data recommendation system based on IPTV application...