Research And Application Of Recommendation Algorithm Based On Graph Neural Network

With the explosion of online information,recommendation algorithms play an increasingly important role in alleviating information overload.Recommendation algorithms aim to use the available data and information to provide users with accurate and diverse services.In addition,research on graph neural network techniques has gradually emerged in recent years,which can learn node information and topological structure information in graph data and has been widely used in various fields.Most of the information used in recommendation algorithms has graph structure,so graph neural networks are very suitable for recommendation algorithms.This thesis will research on social recommendation algorithms and knowledge graph recommendation algorithms based on graph neural networks.For social recommendation,the main data involved are user-product interaction data and user-user social relationship data.Existing social recommendation algorithms do not sufficiently model the internal factors of user-product and user-user interaction,and most algorithms do not differentiate between different users or products.To address these problems,this thesis proposes the Graph Attention Memory Social Network algorithm,which consists of the attention memory module,the attention aggregation module and the model prediction module.The attention memory module captures the multi-faceted influence of users on users and the inherent multi-faceted preferences of users on products,the attention aggregation module aggregates information of neighborhood nodes based on graph neural network and uses attention mechanism to distinguish the different impacts of users or products with different strong and weak relationships on target users,and finally the model prediction module completes the prediction of users' ratings on products.The experimental results show that the Graph Attention Memory Social Network algorithm reduces the prediction error by about 2% compared with mainstream social recommendation algorithms,thus providing some improvement to the recommendation performance.For the knowledge graph recommendation,the knowledge graph data involved has various node types and multiple relationship connection structures.The existing recommendation algorithms have insufficient information extraction capabilities for such heterogeneous graph data with complex structures,and cannot deeply explore the rich knowledge information in the knowledge graph.To address these issues,this thesis proposes the Knowledge Graph Enhanced Bilinear Graph Convolutional Network based on graph neural networks.The algorithm uses a finely designed information propagation layer to aggregate the information of neighbor nodes of target nodes in graph data,captures signals of interactions between neighboring nodes of the target nodes,enhances the embedding information representation of product nodes using the information of multi-relational attribute entity nodes of product nodes in the knowledge graph,and stacks multi-layer information propagation layers to aggregate the information of multi-hop neighbor nodes and explore higher-order connectivity information.The experimental results show that the Knowledge Graph Enhanced Bilinear Graph Convolutional Network has a 2-9%improvement over the mainstream knowledge graph recommendation algorithms in a variety of recommendation metrics.The recommendation algorithm based on graph neural network proposed in this thesis can be applied to video recommendation,intelligent marketing,smart tourism and other application scenarios to provide accurate and personalized recommendation services for users.