Research On Key Models And Algorithms Of Graph Neural Network Session Recommendation System

With the rapid development of the Internet,various network applications have emerged,providing people with abundant and diverse information services.However,the massive amount of information has brought troubles to users,and information overload has become a major challenge facing the current Internet.Recommender systems have emerged to help users quickly find the content they are interested in from the information flow.Traditional recommender systems usually rely on users’ historical behavior data,which often involves user privacy issues and is difficult to obtain in many scenarios.Therefore,session-based recommender systems have received extensive research attention.Session-based recommender systems recommend the next item of interest to users by capturing real-time interaction information between users and the system.Session-based recommender systems not only effectively alleviate the impact of information overload but also avoid user privacy issues by not relying on user historical behavior data.This thesis deeply analyzes and studies current session-based recommender systems,focusing on the session recommender system based on graph neural networks,and summarizes the general modeling steps of the session recommender algorithm based on graph neural networks.Corresponding solutions are proposed to address the shortcomings of some current algorithms.This thesis proposes a session-based recommender model based on a multi-channel self-attention graph neural network.To address the low efficiency of utilizing information in the session graph constructed in the session recommender algorithm,a new session graph construction method and a multi-channel graph neural network are designed for the session graph.Meanwhile,to better learn the global interest preference,attention mechanisms are introduced and multi-layer self-attention modules are designed.Experiments on standard datasets show that the proposed model outperforms baseline algorithms significantly.This thesis proposes a session-based recommender model based on a star topology graph neural network.Considering that there may be potential associations between nodes not directly connected in the session graph,a star topology session graph is designed and a star topology graph neural network is designed for the session graph.Meanwhile,considering that the same type of element in the session may have different effects on global preference coding at different positions in the sequence,position embedding technology is introduced to distinguish the position of the item.Experiments on standard datasets demonstrate the effectiveness of the star topology graph neural network session recommender algorithm.