Edge Information Based Conversational Recommendation Of Gated Graph Neural Network

Recommender systems have emerged as a prominent research area in domains such as e-commerce,online education,and video streaming.However,the difficulty of generating high-quality recommendations arises when the amount of user interaction data is either too sparse or too abundant,leading to the challenging problems of cold-start and redundancy.Prior research has tended to overlook the importance of marginal information on users,potential links between group users,and user session data,resulting in less effective solutions to these recommendation problems.To address this gap,we proposes a novel approach that leverages marginal information,groups,and sessions to study the aforementioned issues,utilizing user online behavior data from the China Internet Network Information Center and a dataset of computer course evaluation exercises.The main research objectives are as follows:Firstly,We propose an edge information-based cold-start algorithm for session-based recommendation systems,called E-CGNN.Specifically,we hot-encode user-item attribute information and improve the autoencoder to reconstruct the attribute information for cold-start users,resulting in user node embedding vectors.We introduce session-based recommendation training methods to obtain predicted ratings of users for websites,based on which we make accurate recommendations for new users,thus implementing the E-CGNN algorithm.Our experiments show that the E-CGNN algorithm outperforms the cold-start adaptive algorithm(PNMTA)by reducing the root mean square error and mean absolute error by 2% and 5%,respectively,indicating improved recommendation performance.Secondly,A session-based recommendation algorithm,E-SGNN,is proposed based on marginal information.Initially,similar users are clustered based on marginal information and divided into different user groups.A session-user-site graph is constructed,and data features in the graph are extracted.Then,GGNN is used to reset and update the user’s historical and current behavior feature vectors.Secondly,a self-attention mechanism is introduced to adjust the proportion of the user’s current preference and historical preference.Finally,a ranking score is obtained using linear transformation and a softmax classifier to implement the E-SGNN algorithm.The experiments show that the E-SGNN algorithm improves the recall rate and average reciprocal rank by 3.13% and 1.24%,respectively,compared to the FGNN-SG algorithm for the cross-session recommendation,indicating an improvement in recommendation performance.Thirdly,We propose a session-based personalized exercise recommendation algorithm,E-SGNN-PER,based on edge information.By redefining the meanings of session,user,and item in the E-SGNN algorithm for personalized exercise recommendation,the session part is replaced with a set of knowledge points,resulting in the E-SGNN-PER algorithm.Experimental results on the Exercise dataset show that the E-SGNN-PER algorithm outperforms the multi-task feature learning recommendation algorithm based on knowledge graph enhancement(MKR)in terms of recall and precision by 2.12% and 1.2%,respectively.