Research And Implementation Of Personalized Recommendation Based On Variational Autoencoder

In recent years,with the development of information technology,the amount and production speed of online data are growing rapidly.Although these network data bring convenience to users,they also cause a problem,"information overload".With the rise of personalized recommendation technology,users can extract effective information quickly from larger amount of complex data.Many researchers have successfully applied deep learning algorithm models to recommendation algorithms.which can greatly improve the performance of the recommendation.However,recommendation algorithm still has many challenges,such as data sparseness,cold start,and accuracy.Variational autoencoder model(VAE)can excavate the relationship between the user and the project well,and has a good expansibility,which is widely used in recommendation algorithms.The current recommendation algorithms research mostly consider the user's social relationships,and rarely consider the social relationships between items.How to effectively combine auxiliary information such as social relationships between items is still a challenge for the recommendation algorithm.When the VAE model trains the data,it may ignore the hidden variables,which leads to the weak correlation between the hidden variables and the observed data,which affects the recommendation results.Aiming at these two problems,this thesis studies and improves the application of the VAE model in the recommendation algorithm.The main works are as follows:1.A recommendation algorithm for integrating item social regularization and VAE model was proposed.The related theories of the VAE model and its application in the recommendation algorithm are studied,Mining the social relationships between items,integrating the item lable matrix,item content information and the social relationships between items into the VAE model.Effectively use the auxiliary information in the item relationships to alleviate the problems of cold start of the item and too sparse data.The implicit feature value and potential content variables of the user-item are learned by Bayesian generative model,and the deep feature representation ability is used to alleviate the inefficiency and sparsity of matrix decomposition in traditional collaborative filtering algorithms.Experimental results show that the improved algorithm is more accurate in predicting results.2.A recommendation algorithm for maximizing mutual information and VAE model was proposed.In the training process of the VAE model,the excessive regularization ignores the consideration of the observation data,which makes the correlation between the latent variable and the input data not strong.This thesis uses the mutual information maximization VAE to construct the model.In the training process,the mutual information maximization item is introduced,a new generalized evidence lower bound object is derived,and parameters are introduced to balance reconstruction loss,regularization loss and mutual information loss.Maximizing the mutual information between the original input data in the model and the learned latent variable representation,using effectively latent variables to get a better item vector representation,and enhancing the correlation between latent variables and input data in the model.The experimental results show that the improved algorithm is more effective in prediction.3.Designing and implementing an article recommendation system.Applying the improved recommendation algorithm proposed above to the article recommendation system,and recommend articles that may be of interest to users.