Research And Application Of Recommendation Algorithm Based On Knowledge Graph And Graph Neural Network

The advantage of knowledge graph(KG)assisted recommendation lies in that knowledge graph can integrate various information,such as item attributes and user data,effectively alleviating the data sparsity problem commonly found in traditional recommendation methods;As an effective deep learning method for extracting features from graph data,graph neural network,combined with knowledge graphs with graph structure features,can effectively improve the effectiveness and efficiency of intelligent recommendation algorithms.This is the starting point of this thesis.In summary,this thesis aims to improve recommendation accuracy and diversity,focusing on issues such as node neighborhood quality,sample noise,and node oversmooth.A knowledge graph recommendation algorithm based on two typical graph neural network is proposed and applied to the personalized film recommendation prototype system.The specific research work and achievements of this thesis are as follows:(1)Research on recommendation optimization algorithms according to KG and graph convolutional network(GCN).At present,the combination of KG and GCN has formed the general framework KGCN.Based on this,an important neighbor sampling method is designed to address the problem of difficulty in ensuring the quality of target node neighborhoods in constructing KGCN.This method utilizes node centrality to achieve priority sampling and node hiding,thereby improving the quality of node neighborhoods.In addition,in regard to the common sample noise problem in the recommended data set,A Denoising Auto Encoder is designed at this model input layer to draw more robust features of users and items and prevent overfitting of the model.The recall rate and accuracy are selected as evaluation indicators and verifies the effectiveness of the proposed optimization algorithm through comparative experiments.(2)Research on recommendation optimization algorithm according to KG and graph attention network(GAT).As the number of layers increases in the learning process of GAT,there is a tendency for excessive node smoothing,which affects the diversity of recommendations.For this reason,based on the general recommendation model KGAT of KG and GAT,this thesis designs the collaborative knowledge graph of users and items,and then uses the neighborhood representation of nodes in the model feature propagation layer to build an ordered sequence.With the help of GRU,which has strong sequence modeling ability,it controls the information flow between layers to suppress information interference.Simultaneously combining residual networks,residual preference learning connections are used throughout the entire feature propagation layer to capture user preferences and improve recommendation performance.This thesis compares it with other baseline models on three open datasets,and experimental results imply that the algorithm validly raises the diversity of recommendation.(3)Design and realization of the film personalized recommendation prototype system.In order to apply the designed optimization algorithm to practical applications,this thesis designs and implements a personalized movie recommendation prototype system that integrates two recommendation algorithms.The backend business module of the recommendation system is developed using the SpringBoot framework,and the main body of the front-end display page is implemented using Angular JS.The effectiveness of the system operation reflects the effectiveness of the system design and the accuracy and diversity of recommendation algorithms.