Research On Recommendation Algorithm Based On Graph Neural Network In Heterogeneous Graphs

With the rapid development of information technology,heterogeneous data gradually becomes the main form of information representation,and many heterogeneous data can be naturally modeled as heterogeneous graphs.Recommendation systems,as an important topic in the field of heterogeneous graph research,are widely used in many practical recommendation scenarios.Graph neural networks can naturally integrate node information in heterogeneous graphs and have powerful graph learning capability to serve heterogeneous graph recommendation.However,existing graph neural network-based recommendations usually focus on the node representations of users and items,ignoring the rich interaction information between users and items.Rich interactions can be learned using meta-path information,but heavily rely on reachable explicit paths,and anomalous links in graphs can also affect the accuracy of recommendations.The main work of this paper is as follows:1st.To address the problem that existing graph neural network-based recommendations ignore rich interaction information,a recommendation model based on bi-order metapaths and graph neural networks is proposed.The model focuses on the fusion of users,items,and metapaths,learning node representations of users and items using graph attention networks,and learning interaction information in meta-paths for final recommendations.The model also designs a novel meta-path selection method that considers bi-order meta-paths with special semantic information and extracts the interaction information in the meta-paths by convolution.2nd.To address the problem of sparse connection of explicit reachable paths,a recommendation model based on neighborhood interaction and graph neural network and convolutional optimization are proposed.The model uses a neighborhood-based meta-path sampling method that does not depend on the reachable explicit paths,and uses the Fourier transform to optimize the convolution calculation of the neighborhood interaction after obtaining the meta-path neighborhoods of users and items,which reduces the time complexity of the model.Compared with the best method,the average accuracy of the model is improved by 1.76%～5.95% on average.3rd.To address the problem of anomalous links exacerbating data sparsity,a graph neural network social recommendation model with data augmentation is proposed.The model obtains the existence probability of links through a link predictor,then adds missing links and removes noisy links according to the probability,and incorporates social information for recommendation on the basis of the new graph obtained.The data augmentation improves the accuracy of the proposed model in this paper by 4.0%.Focusing on the three problems of insufficient recommendation aggregation information,path sparsity affecting performance,and anomalous links exacerbating data sparsity,this paper fuses the information of users,items,and meta-paths,and uses Fourier convolutional superior neighborhood interaction,while reducing anomalous links and incorporating social information.The performance is excellent in the recommendation task of heterogeneous graphs.