Research On Recommendation Algorithm Integrating Social Graph

With the continuous progress of human civilization and the rapid development of science and technology,network technology and big data technology have been rapidly developed and widely used.Businesses and users in the vast sea of data and information,how to quickly find the information they need,it is particularly important.Nowadays,in e-commerce and social media platforms,the application of recommendation system can filter out a large number of irrelevant messages and quickly and accurately locate the useful information needed by users,which plays an important role in users' quick search for useful information and easing information overload.Networks are ubiquitous in the real world,representing objects and relationships in different domains.Both traditional machine learning and earlier deep learning have to some extent ignored the structure between data--graphs.In order to solve these problems effectively and improve the accuracy of recommendation,this paper analyzes the causes and solutions of these problems,researches and discusses the graph neural network integrating social relations,and proposes a new recommendation system algorithm which is formed by the combination of graph neural network and attention mechanism.Recommendation systems need data support when recommending products.In order to solve the problem of lack of data--the problem of cold start,we try to use auxiliary information--social network.Since cold-initiated users are more dependent on social networks than users with higher ratings,the effect of using trust propagation on cold-initiated users becomes more important.In addition,in many real life social rating networks,a large proportion of users do not express any ratings,they just participate in the social network.Therefore,only using observed scores is disqualifying.The social influence is brought into the recommendation system to make use of the interaction and relationship between users to assist the recommendation system to recommend users.The thesis mainly completes the following three aspects of work.Firstly,the social aggregation method of graph neural network for social recommendation is modified,and graph attention network is used.At the same time,deep and crossover networks are used in the fusion user modeling,in which the interaction strength and memory ability between features are increased by adding crossover layers.Secondly,for social relationships,aggregating information from inconsistent social neighbors affects the ability of graph neural networks to aggregate beneficial information to make recommendations.Social inconsistencies can be divided into two types: first,there are completely inconsistent items between users connected in the social graph.Second,users connected in the social graph have opposite opinions about the same project.In order to solve the problem of social incongruence,this paper analyzes the information from the incongruent social neighbors,and proposes a combination of query layer and relational attention module to distinguish the influence of different friends.Thirdly,in order to demonstrate the accuracy and validity of the model,the ablation experiment was used to verify the model.