Study On Social Recommendation Systems By Integrating Matrix Factorization With Graph Embedding

With the rapid development of artificial intelligence technology,the boundary between human society and the information world has become increasingly blurred.Under this background,the types and quantity of data are increasing day by day,and how to select valuable information quickly and accurately from the vast amount of big data has become one of the main problems faced by the academia and industry together.Recommendation system,as an effective complementary means of traditional information retrieval,makes full use of the content characteristics of users and items and the interactive data between them to automatically filter useless information,in order to combat the problem of information overload and achieve the equilibrium between information producers and consumers.Collaborative filtering,especially matrix factorization,is one of the core technologies behind the recommendation system.It skillfully uses the crowd sensing idea to realize personalized recommendation.Due to its stable prediction performance and flexible scalability,collaborative filtering has always been a research hotspot in academia and industry.However,its performance is often limited by the problems of data sparsity and cold start.At present,the mainstream solution is to utilize social network among users to make up for user-item interaction data.However,due to the complexity of social network data,most of the existing social recommendation algorithms are heuristic,which fail to fully mine useful information in social networks to assist recommendation tasks.Fortunately,the rise of graph embedding technology provides a new idea for the study of social recommendation,which is committed to embedding high-dimensional sparse social information into low-dimensional dense vector space,while maximizing the structural information of the original network.In view of this,this paper takes the social recommendation system as the research object,focuses on exploring the effective fusion scheme of matrix factorization and graph embedding technology,proposes two fusion paradigms of ordinal learning and joint learning.Now the main work is listed as follows:Firstly,aiming at the inherent data sparsity problem of collaborative filtering technology,an algorithm framework for ordinal learning of matrix factorization and neural graph embedding technology,SoTriCF,is proposed.Three classical collaborative filtering methods are integrated to capture the multiple similarities between users and items.Then the low-dimensional user representations obtained by graph embedding technology to process social information are integrated into recommendation task in turn,so as to alleviate the twofold sparsity of rating data and social data;In addition,a solution based on the latent feature mapping mechanism is proposed to alleviate the cold start problem.Besides,in order to further achieve the effective integration of recommendation task and social network analysis task,an integrated method for joint learning of matrix factorization and neural graph embedding model,NGE-MF,is proposed to realize the two-way interaction and collaborative optimization between the rating prediction and user network embedding tasks.Matrix factorization technology can learn the user's behavior habits of the items,and neural graph embedding technology can fully capture the user's social characteristics.Through the joint learning of the two parts,the learned user's hidden features can be optimized in a unified framework,so as to obtain more realistic hidden features.Finally,through a large number of comparative experiments and parameter analysis on three real-world datasets,the experimental results show that the two combination paradigms of social integration schemes have achieved better recommendation performance,and can effectively alleviate the problems of data sparsity and cold start,which fully demonstrate the rationality and effectiveness of the proposed algorithm.