Research On Session-Based Recommendation Algorithm Based On Graph Neural Network

With the rapid popularization and development of the Internet,people can access massive information conveniently.But at the same time,they also suffer from “information overload”,and session-based recommendation can effectively alleviate this problem.The combination of session-based recommendation algorithm and graph neural network technology has achieved a series of results,but they are yet to be improved.For example,time information in session sequence is not being sufficiently used and much of the relationship between sessions still awaits to be explored.In this thesis,the above problems are studied,and several improving methods are proposed.The main work is as follows:A session recommendation algorithm fused with time context information is proposed.The time information in a session is very important for modeling user behavior.For better use of time information,the thesis designed a Time-LSTM based graph neural network which,on the basis of the ordered session graph,uses Time-LSTM as the aggregation strategy to facilitate effective modeling of user behavior with the time information of the session.An attention mechanism based on time information is further designed to dynamically calculate the different importance of different session nodes,and then obtain a better representation of user features.A session recommendation algorithm combining global session information is proposed.For better use of the relationship between anonymous sessions,a module for extracting global session information is firstly designed.Construct a global session graph based on the frequency information of global nodes and the co-occurrence magnitude between nodes,and use graph neural network to obtain a feature representation of session nodes containing global information.Then a new recommendation model is obtained by combining the local information extraction module of the session.This method not only takes the node relationship within the current session into consideration,but also makes use of the global information between sessions to model user behavior.As a result,a more accurate representation of user features can be obtained by combining the two parts of informationIn this thesis,the proposed model is fully tested on three public data sets,Yoochoose,Tmall and Nowplaying.Experimental results show that the two models proposed in this paper have an improvement of up to 4.70%on P@20 and 2.58% on MRR@20,which verifies that the proposed methods can effectively improve the recommendation effect.