Research On Recommendation Algorithm Based On Variational Autoencoder

With the explosive growth of Internet information,how to excavate and analyze the information has become a hot and difficult problem.The emergence of the recommendation system has effectively alleviated the problem of information redundancy.The essence of the recommender system is to mine the user's preference pattern through the user's historical behavior information and complete the recommendation according to the requirements of different scenes and tasks.In the traditional recommendation methods,the latent factor model can effectively mine the latent representation of users and items according to the user behavior history.The deep model expands the linear operation of the traditional model and improves the expressive ability of the model.However,there are still some problems in the research in the field of recommender systems.Due to the existence of massive data,for a single user,the interactive items usually only occupy a very rare part of the massive items.And due to the limited content of interactive information,it is difficult to obtain the priority relationship between the user's preferred items.Besides,user preferences will develop over time,and it is difficult for general recommendation systems to capture the direction of changes in user interests.Variational autoencoders(VAEs)have been widely used in the field of recommender systems in recent years.As a nonlinear probabilistic generative model,the variational autoencoders have the advantage of breaking the limited modeling ability of linear models and combining the flexibility of the distribution definition of the generative model with the ability of deep learning to learn nonlinear features.Based on the variational autoencoder model,we propose a ranking model and a sequential model,aiming at the problems of data sparsity,recommendation accuracy,and recommendation changes over time in the recommender system.The main contributions are as follows:(1)Aiming at the problem that some deep generative recommendation methods cannot model implicit feedback user preference data to generate a personalized list-wise ranking recommendation list,we enforce a list-wise ranking strategy to VAE with the aid of multinomial likelihood.This model can simultaneously generate the point-wise implicit feedback data and create the list-wise ranking list for each user.To seamlessly combine ranking loss with variational autoencoder loss,the mean reciprocal rank(MRR)and normalized cumulative loss gain(NDCG)are adopted here and approximated with a smoothed function.A series of experiments on three real-world datasets have been conducted.We show that maximizing the ranking loss will cause as many relevant items to appear at the top of the predicted recommendation list as possible.The experimental results demonstrated that the proposed deep generative ranking method outperforms several state-of-the-art methods in a ranking estimation task.(2)Given the problem that some deep generative recommendation methods do not use the sequential information in the implicit feedback user preference data for personalized recommendation,we combine the variational autoencoder with multinomial likelihood and the sequential feature generative module to complete the recommendation task.The model extends the variational autoencoder by utilizing the rich information in the user's historical preference records.The model provides a modeling scheme for generating observation data according to the time-sequential relationship and introduces a sequential feature generation module,which is used to capture the user's interest evolution pattern related to the relevant item.Besides,the extended form of the model is given,which extends the model prediction from the behavior of the user at the next time step to the behavior of the next k time steps.A series of experiments on two real-world datasets have been conducted.We show that handling temporal information is crucial for improving the accuracy of the variational autoencoder.The experimental results demonstrated that the proposed deep generative sequential method outperforms several state-of-the-art methods in the time-sequential recommendation task.