Research And Application Of Debias Recommender System Based On Deep Representation Learning

With the development of deep representation learning technology,the recommendation system can more intelligently generate the representation vectors of users and items,and recommend users the items which they are interested in from massive information.However,due to a series of reasons such as the interactive feedback mechanism between users and items,the negative sample sampling method,and the long-tail distribution of source data,the recommendation systems suffer different types of bias which causes varying degrees of impact on the improvement of model performance.This thesis mainly explores the bias problem in the recommendation system of deep representation learning.The main contents of this thesis are as follows:(1)A Debiased Recommendation Algorithm Based on Pseudo-Label Generation.In the process of data sampling,the recommendation model usually marks the non-interaction samples as negative labels,ignoring the potential positive data in the non-interaction sample pairs.Therefore,this thesis establish a pseudo-label learning framework suitable for recommendation systems,adopting Monte-Carlo dropout and multi-model training strategies to improve the credibility of labels,and introduce inverse propensity scoring to balance the impact of items with different exposure ratios,thereby effectively alleviating bias and improving performance.(2)Self-supervised Learning Debiased Recommendation Algorithm Based on User Interest.Users and items in the recommendation source data follow long-tail distribution.In order to alleviate the popular bias in the recommendation system,this thesis proposes a self-supervised learning framework based on user interest degree,which obtains user interest representation embedding through history sequence.Besides,a two-stage feature mask based on mutual information and sequence random mask are used for model perturbation.The contrastive loss is calculated and added to the loss function for model training,which alleviates the popular bias and improves the prediction performance.(3)A Debiased Enterprise Recommender System Based on Tax Data.Due to the lack of effective negative sample labels for tax invoice data,the enterprise recommendation scenario based on tax data is a typical industrial scenario with recommendation bias.This study applies the debiased recommendation model to the enterprise recommendation scenario,with an interpretable hybrid recall strategy and a complete multi-cascade recommendation architecture.The thesis focuses on the bias in the recommendation system,alleviates the exposure bias and popular bias of the recommendation system through semi-supervised learning and self-supervised learning,and finally applies the debiased recommendation algorithms to the enterprise recommendation scenario.The research results of this thesis have improved the performance of the recommendation model,alleviated the recommendation bias,and have been deployed in application scenarios,providing innovative improvement ideas and application cases for alleviating the bias of the recommendation system.