Research On Recommendation Algorithm Based On Graph Neural Network And Social Network

Today’s society has a wealth of digital content to amuse itself with,such as e-books,movies,videos and online shopping,but the huge amount of data also leads to the problem of "information overload".The emergence of recommendation system is to solve this problem,and now it is closely related to People’s Daily life.Whether it is buying clothes,watching wonderful You Tube videos,or searching for restaurants in a new city,the back-end recommendation system provides support for these services.In recent years,how to use recommendation system to accurately recommend items that individuals(groups)may be interested in has been the focus of scholars’ research.The combination of attribute graph and graph neural network in personal recommendation algorithm can indeed improve the accuracy of algorithm recommendation,but the combination of graph neural network and attribute graph will embed all attributes of users or items.Users learn items in the same way,which leads to some attribute noise in the recommendation process,and can not make good use of attribute graph for recommendation.In addition,the existing model treats the interaction between users and objects as equivalent,while in reality,users will have different preferences for different attributes of objects.At the same time,with the rapid development of social media,people gradually shift the research direction from personal recommendation algorithm to group recommendation algorithm.Most of the existing group recommendation models adopt heuristic or attention-based preference aggregation strategies to learn the personal preferences of group members and aggregate them into group preferences.However,due to the sparsity of user interaction data,the user characteristics after learning are not complete,and the user interaction in real life is very complex,and the user relationship may be high order.Furthermore,the similarity between groups and the personal preferences of common members are often ignored,and group similarity has great potential to improve group representation learning.In view of the above problems,the main research contributions of this paper are as follows:1.Aiming at the problems of different attribute preferences and noise in attribute graphs,a collaborative filtering recommendation algorithm combining the graph attention network and beneficial features is proposed.Firstly,in the interaction process of user attribute graph and item attribute graph,the customized graph attention algorithm is used to update the interaction information of different weights between attributes.Secondly,the attribute characteristics of users and items are screened by the beneficial feature detection algorithm,and the attribute interaction that is unfavorable to recommendation or has low impact is abandoned.Finally,the feature representation of the user in the user map and the feature representation of the item in the item map are matched to better complete the recommendation.Through the experiments on the real data sets Taobao and Book-crossing,AUC,Logloss and NDCG@10 are used for evaluation.The results show that the model proposed in this paper is superior to other common benchmark models.This proves that the combination of graph neural network and beneficial feature detection attribute graph can improve the accuracy of providing recommendations to a single user.2.Aiming at the problems of sparsity and complexity of user data in groups and similarity among groups,a group recommendation algorithm based on fusion hypergraph convolution and self-supervised cooperative training is proposed.First,in the user-level hypergraph,three channels are used to encode the high-level user relationship in the hypergraph convolutional network,and the enhanced user representation is obtained by aggregating user characteristics learned by multiple channels,which provides a solid foundation for learning group preferences.Secondly,in the group-level hypergraph,all groups are connected as overlapping networks,and the personal preferences of common members of the group are concerned,in which the process of hyperedge embedding can be regarded as the learning of group preferences.To further enhance group representation,self-supervised learning and cooperative training were combined to construct two different graph encoders on the above two hypergraphs,which recursively used different information to generate labeled samples and supervised each other by comparing learning strategies.Compared with discard strategies,the proposed self-supervised cooperative training retained complete information and realized real data enhancement.The proposed HCSC model has been extensively tested on two real world data sets,and the experimental results show the superiority of the proposed HCSC model.