Research On Session-based Recommendation Algorithms Based On Gated Graph Neural Networks

With the rapid development of information technology and the increasing popularity of intelligent terminal devices,the Internet not only provides people with convenience in their daily lives,but also brings serious "information overload" problems with the massive amount of data information it generates.Recommendation systems have been proposed as a type of information filtering technology to solve the problem of "information overload" and are widely used in various Internet-related products.Session-based recommendation systems are one of the current research focuses in the field of recommendation systems.They are suitable for recommendation scenarios that lack user’s personal information and long-term historical behavior data.The recommendation task is accomplished by mining the user’s interests and preferences in the current session.Compared to traditional recommendation systems,sessionbased recommendation systems pay more attention to the changes of users’ current interests and preferences,and can make more accurate and timely recommendations.This study centers on gated graph neural network and attention mechanism,and mainly carries out the following work:(1)Session-based recommendation algorithms based on graph neural network generate the embedding representation of items only by aggregating the information of neighboring nodes when modeling the current session sequence,ignoring the high-order transformation relationship between non-adjacent items.To solve the above problem,a session-based recommendation model based on high-order gated graph neural network(Ho-GNN)is proposed.Ho-GNN constructs the corresponding session graph based on the neighbor transformation relationship of different orders in the current session sequence,and updates and propagates nodes information in the session graph through gated graph neural network and attention mechanism to generate embedding representation of items containing more high-order neighbor information.Adopting a strategy of integrating long-term and short-term interests to generate the session embedding representation that reflects the overall interest preferences of users,and the probability that each candidate item may be interacted with by the user is calculated according to the embedding representation of current session to generate the final recommendation list.(2)A session-based recommendation model with fusing neighbor-hood collaborative information(SR-FNC)is proposed based on Ho-GNN.This model uses item-level collaborative information from other sessions to supplement information of the current session and improve the recommendation performance of the current session.Firstly,SR-FNC searches neighbouring sessions with similar behaviour patterns to the current session,and converges transformation relationships between items in the current session and all neighboring sessions onto a single graph.Scilicet,constructing a neighbouring session graph for the current session.Then,using the gated graph neural network to capture item-level global collaborative information,and generating a global-level embedding representation of the item.Finally,the global collaborative information of the item is incorporated into the modeling of the current session sequence to generate item’s features containing more information to complete the final recommendation task.(3)In this study,we conduct multiple sets of experiments on two real popular e-commerce datasets,Yoochoose and Diginetica,and results of these experiments show that two models proposed in this study,Ho-GNN and SR-FNC,can generate more accurate recommendation results and verify the effectiveness of these models.