| With the rapid development of the Internet and the popularization of smartphones,the amount of information on the Internet has increased rapidly and changed the way people get information.However,the overload of information caused by the large number of information resources hinders people from obtaining accurate recommendations,and the appearance of recommendation systems greatly alleviates this problem.The recommendation system analyzes the data and models the potential behavioral characteristics of users,and then generates personalized recommendations.The data in the recommendation system can be regarded as graph data,the user-project can be regarded as nodes in the graph,and the userproject interaction can be regarded as edges in the graph.Traditional neural networks can’t adapt well to the complex and diverse graph structures in recommended scenarios.In recent years,graph neural networks have great advantages in modeling graph data and mining graph nodes.However,the recommendation system based on graph neural network faces some challenges such as over-smoothing and sparse data.The proposed algorithm alleviates the damage to the recommended performance caused by the above problems.The specific work is as follows:(1)This thesis presents a recommendation algorithm based on infinite layer convolution simulation in dynamic subgraphs.The algorithm constructs dynamic subgraphs by embedding nodes containing graph structure information and simulates infinite layer convolution on subgraphs and original user-item interactions graph respectively.(2)This thesis presents a dynamic two-stage graph convolution recommendation algorithm.Unlike traditional recommendation algorithms,the first stage of this work uses the classic TopK recommendation model to complement the original user-project interaction matrix to reduce the data sparsity.In the second stage,the proposed algorithm based on infinite layer convolution simulation in dynamic subgraphs is used to predict the data in the proposed scenario.The two algorithms presented in this thesis are compared on three real data sets.Experiments show that this method has better performance on recall,F1 and NDCG than the baseline method. |
