A Signed Graph Social Recommender System Based On Variational Autoencoder

The social recommendation system uses the social relations of the social network to obtain network embeddings,uses the network representation learning method to map the information of users and items to a unified space,and calculates the embeddings matching degree to propose recommended items.This paper studies the application of graph neural network algorithm in network representation learning on signed social networks.Based on graph convolution method,linear convolution kernel is used,which is suitable for the sparse characteristics of network node data and the sparsity of connections,and generates embeddings efficiently.In the process of graph convolution,since the goal of this paper is the graph convolution network for the recommendation system,the original graph convolution may not be effective enough.Therefore,a simplified graph convolution network is proposed to apply to the recommendation task,and the symbol graph convolution in this paper is simplified,so as to make the whole network structure more suitable for the data research of symbolic social networks.The probabilistic distribution randomness is introduced into the variational autoencoder that generates node embedding,which makes the model representation ability exceed the linear model representation ability.This variational autoencoder adopts the principle and technology of variational inference,combined with the symbolic network embedding algorithm,and aims to improve the accuracy of recommendation tasks in social networks.In the framework of generative embedded variational autoencoder,two encoders with different distributions are proposed for symbolic network data,which are convenient to calculate and improve the recommendation effect respectively.Both encoders are used for social network recommendations,embedding symbolic networks at a small computational cost.In order to verify the effect of the research content in this paper on the social recommendation of symbolic network,two publicly available symbolic network data sets,Epinions and Slashdot,were used to conduct the experiment,and two error indicators,MAE and RMSE,were used to evaluate the model effect.Compared with other symbolic network representation algorithms,the results show that the recommendation effect of the proposed method is better than other models,which demonstrates the effectiveness of the proposed algorithm.