Application Of Variational Autoencoder In Recommendation System

With the development of the internet,more and more information is flooding into people’s daily lives.To help people better select information of interest,recommendation algorithms have become a research hotspot in recent years.Recommendation algorithms can predict user preference by analyzing their historical behavior and identifying potential relationships between users and items.Deep neural networks can uncover deep behavioral patterns between users and items,while variational autoencoders can effectively discover nonlinear relationships under linear assumptions and perform well in solving recommendation ranking problems.However,current variational autoencoders still have some limitations that restrict their recommendation performance.This project will focus on these core issues:(1)Currently,recommendation datasets are typically sparse,and traditional variational autoencoders have limited processing capabilities when faced with overly sparse data.This article attempts to address this issue by leveraging the binary nature of recommendation datasets and incorporating graph relationships to construct a bidirectional conditional distribution variational autoencoder.Experiments show that this new model has a significant effect on improving the recommendation performance of sparse datasets compared to similar models.(2)While benefiting from the role of variational autoencoders in recommendation algorithms,the common problem of posterior collapse inevitably arises,which means that the variational encoder cannot learn anything and cannot make accurate recommendations.To alleviate this problem,this article utilizes comment information for encoding and constructs prior constraints to enrich the prior distribution,thereby improving the model’s learning ability.Experiments show that this method further improves the new model based on the original model.(3)Finally,it was found in the experiments that datasets with graph relationships are very limited,which brought some difficulties to the experiments.To facilitate subsequent research,this article provides graph relationship auxiliary information for a series of datasets and conducts experiments on several classical models to provide a baseline.