Collaborative Deep Bayesian Recommender System Under Random Missing Data

In recent years,with the popularity of the Internet and the advent of the ”streaming media” era,recommendation system technology has received extensive attention and research from major Internet companies.Especially used by big companies or apps such as Tiktok,Kuaishou,instagram,and Netflix,they continuously push ”short videos” or information to various users every day.However,in the face of the massive data generated by the Internet every day,how to effectively and high-quality push information is an important issue faced by major companies,and it is also a popular research topic for countless researchers.In many recommender systems,matrix factorization is one of the widely used collaborative filtering methods in recommender systems.Its ideas and mathematical principles are simple and easy to operate,and are favored by many researchers.However,most of the existing studies assume that the data are completely observed and there are no missing cases.In practical applications,people often encounter the situation of missing data.In public datasets like Movie Lens movie ratings,the missing data is sometimes as high as 95% or more.If the relevant information of the missing data is not considered,the recommendation will be invalid.Although this issue has attracted extensive attention of many experts and scholars in recent years,there are still problems such as the need for a hypothesis mechanism to deal with missing data,and the poor explanatory power of the model.Therefore,in order to solve these problems,from the perspective of matrix decomposition,this paper proposes a method of interpolation while learning by using the feature information by using the neural network autoencoder to extract item feature information and user feature information.This paper assumes that the linear model method is more simple,direct and efficient,and the method of interpolation while learning is efficient and interpretable.The main innovations of this paper include: 1.Considering the problem of missing data in recommender system modeling has both practical significance for the model and improves the recommendation accuracy.2.Using item and user feature information to impute the latent variables while learning,which alleviates the impact of missing data on the effect of the recommendation system,and improves the interpretability of the recommendation results.3.This paper simulates the performance of the model under different missing rates,and demonstrates the effect of different missing rates on the model.The following three important conclusions are obtained:1.Considering missing data is helpful to improve the accuracy of the model,and ignoring the problem of missing data does not conform to the actual recommendation scenario and is not conducive to improving the model effect;2.Using the information of items and users themselves The feature extraction and imputation of latent variables have good interpretability;3.Through the simulation study,it can be found that different missing rates have a significant impact on the model effect,which has guiding significance in the future modeling.