Research On Recommendation Algorithm Based On Graph Neural Network

Recommendation system has become an effective way to the solve the information overload problem.Its main goal is to provide valuable and directional content for users according to user preferences.The most important thing in a recommendation system is the recommendation algorithm,which determines the popularity and value of the recommendation system.Collaborative filtering in traditional recommendation algorithm is the most mature recommendation algorithm,but still struggles with the problem of data sparsity.Although the recommendation algorithm based on deep learning can indeed combine auxiliary information to solve this problem,it is difficult to deal with graph structure data directly.In order to address thess issues,in this thesis,the recommendation algorithm based on the graph neural network is studied and the recommendation model is optimized.In view of the fact that the graph auto-encoder can share convolution parameters,and has the characteristics of introducing node features to store auxiliary information.This thesis investigates the problem of predicting movie rating in the recommendation system from the direction of graph auto-encoder,not only considering the features of the node itself but also considering the features of the surrounding neighbors.The original data is preprocessed.Then,the feature of user-item bipartite graph is extracted by the graph convolution neural network,and the model merges user and movie feature information.At the same time,the idea of R-GCN aggregation is introduced,and the decoder reconstructs the rating matrix.The rating prediction model has better recommendation performance for sparse rating data,and expertises analyzing different entities and their relationships in a recommendation system.In order to further improve the performance and effectiveness of the model,the loss function,optimization function and Dropout ratio were selected based on the rating prediction model to optimize the model.The errors generated by the analysis model was analyzed from the perspective of train set and test set.Since the Adam optimization algorithm has the problems of poor convergence and generalization ability,this thesis selected the AdamW optimization algorithm which has the characteristics of fast convergence,strong generalization ability,fast operation speed and suitable for sparse data,and the Cross Entropy Loss function which is suitable for multi-classification tasks.This thesis proposed a graph auto-encoder rating prediction model based on the AdamW optimization algorithm.In order to determine the optimal optimization model,this thesis selected and set the parameters of the AdamW optimization algorithm and the Dropout ratio,and also verified the convergence of the optimization model.In this thesis,Movie Lens dataset is selected as the experimental object.The validity of the movie rating prediction model based on graph auto-encoder was verified,and the RMSE of the model reached 0.917.The errors generated by the comparison of the baseline models were reduced,which showed that the graph neural network recommendation algorithm can improve the prediction accuracy to a certain extent.The optimization model was compared with other models,and the error of verification optimization model was lower than those of comparison models.The optimization model reduced the RMSE by 0.76% over the movie rating prediction model based on graph auto-encoder.It was confirmed that the optimization model had better rating prediction ability for sparse data,which improved the performance and stability of the model to a certain extent.