Research On Recommendation Method Based On Wasserstein Graph Auto-encoder

Recommender systems are software tools and algorithms designed to help users find items of interest to them by predicting their preferences or ratings for items.The most popular recommendation method currently is the collaborative filtering recommendation method,which utilizes the interaction of users and items to build a model for recommendation.Therefore,how to improve the recommendation quality of collaborative filtering recommendation methods is the focus of research in the field of recommendation.In recent years,domestic and foreign scholars have proposed many models to improve the recommendation quality of collaborative filtering recommendation methods,and the model based on graph neural network is an important part of it.Recommendation prediction problem can be mapped to link prediction problem in graph-structured data,and graph neural network has natural advantages in link prediction.As a graph neural network,graph auto-encoders are often used for link prediction tasks due to their strong generative ability.However,most of the existing graph auto-encoders improve on the variational graph auto-encoder model.There are two problems in the variational graph auto-encoder model: a variational posterior distribution is used for each node to force a matching prior distribution,and different variational posterior distributions may cross,which is not conducive to graph reconstruction;Gaussian distribution It may cause that the variational posterior value is always equal to the prior distribution value,and the model will fall into a "local optimum",so that the latent variables are treated as noise,which seriously affects the recommendation quality.This paper proposes a new graph auto-encoder—Wasserstein graph auto-encoder.On this basis,a collaborative filtering recommendation method based on Wasserstein graph autoencoder is constructed to obtain better recommendation quality.The contributions of this paper are as follows:1.In this paper,the Wasserstein distance is effectively applied to the graph auto-encoder model,and the Wasserstein Graph Auto-Encoder(WGAE)model is constructed.The WGAE model uses a continuous mixture distribution to match the prior distribution,and the latent variables of different nodes are far away from each other in the mixture distribution,which is conducive to the reconstruction of the graph,which is more conducive to improving the quality of recommendation in collaborative filtering tasks.2.A new distribution is used in the WGAE model-the von Mises-Fisher(v MF)distribution.The density parameter of v MF is regarded as a hyper-parameter in the WGAE model,which structurally prevents falling into a "local optimum".At the same time,the v MF distribution can also enable the model to learn a richer topology during the nonlinear dimensionality reduction process.3.The WGAE model is applied to the collaborative filtering task,and a collaborative filtering method WGAE-CF based on WGAE is constructed.In this paper,the proposed WGAE model is compared in the link prediction task,and the WGAE-CF method is compared in the collaborative filtering task.Experiments show that the WGAE model can improve the link prediction accuracy,and its performance is higher than the baseline model on the three citation network datasets.Compared with the baseline model,the WGAE-CF method has a lower RMSE value in the evaluation index in the collaborative filtering task.At the same time,the cold start problem has also been improved to some extent.