Research On Recommendation Algorithm Based On Social Networks And Graph Embedding

The personalized recommendation aims to identify user preferences from massive user data,support the personalized content recommendation,advertisement,and marketing management.The existing personalized recommendation methods mainly model the user and item attributes,and the historical interaction information between the users and items.These methods suffer from data sparsity and cold start.With the development of Internet technology and the popularity of social media,introducing massive social network data as auxiliary information into recommendation systems has attracted widespread attention.The user-item interaction network and social network have natural graph structures,so how to process the graph data and extract specific features have become the primary task of the recommendation problem.The graph embedding technology can transform the graph data into a low-dimensional vector representation,and the generated embeddings can preserve the topological structure and potential features of the graph to support the downstream recommendation task.Taking social network data as auxiliary information and effectively representing complex heterogeneous information networks constructed by social networks and user-item interaction networks can partially alleviate data sparsity and cold-start problems.Therefore,the research of recommendation algorithm based on social networks and graph embedding has important theoretical and practical significance.Recent years,the researches on social recommendation mainly focus on modeling explicit and implicit relations in social networks,ignoring the special phenomenon that the importance of high-order implicit relations to different users may vary greatly.In addition,most existing graph embedding methods model social networks and user-item interaction networks as static graphs,while real-world information networks evolve.How to effectively represent complex dynamic graphs have become a challenge.Therefore,this research conducted the following studies:(1)This research propose a social recommendation model for static information networks.Our model consists of three components.Firstly,the unreliable relations between users are filtered by similarity calculation,the potential reliable relations between users are identified to reduce the negative impact of noise data on the model performance,and partially alleviate the data sparsity and cold start problems.Secondly,an adaptive random walk algorithm is designed based on the node centrality to customize the personalized walk strategy for users.Constructing the corresponding higher-order implicit relations,and then complete the reconstruction of the graph.Finally,the graph convolution network is introduced to aggregate neighbor nodes information,which flexibly assign weights to neighbor nodes to generate more accurate user embedding,and complete higher-order implicit relations modeling to further alleviate data sparsity and cold-start problems to support the social recommendation.(2)This research propose a social recommendation model for dynamic information networks.The model consists of three components:embedding generation,interactive information aggregation and relation information aggregation.Firstly,considering the user-item interaction information and scoring information simultaneously,the user and item embeddings are generated.Secondly,users and items are modeled from the perspectives of temporal and static structure.Finally,aggregating user relation information and item relation information to model user and item,and then supporting the downstream social recommendation.(3)This research perform extensive experiments on five real datasets covering the task of social recommendation,empirically verifying the effectiveness and rationality of proposed models.