Research On The Recommendation Method Of Removing Popularity Bias

Recommendation system,natural language processing and computer vision are known as the three major directions of machine learning.With the development of machine learning and deep learning,recommendation system has also achieved great development.Recommendation system has attracted the attention of many researchers both in academia and industry.In daily life,it is not difficult to directly apply the recommendation system in common applications such as e-commerce shopping,social platforms,video platforms and so on.However,due to the data deviation,which is also an important problem of missing random data(MAR)in the statistical field,the recommendation system always has the popularity deviation.This deviation will lead to the recommendation system always preferring popular commodities with a large number of recommended interactions,but this leads to the convergence of the recommendation system and seriously affects the user experience.This article first uses T-SNE dimensionality reduction commodity embedding to verify the prevalence of popularity bias.To address the issue of popularity bias,this article proposes two solutions.The main goal of the first solution is to construct unbiased commodity embedding.Based on experiments,we find that popularity bias can be used as an intrinsic attribute of a commodity.In Chapter 3,we propose a method for decoupling popularity bias,In the dataset,each product is divided into popular and non popular product groups based on total interaction quantification,and fitted using a clustering algorithm to achieve the goal of decoupling popularity and removing bias.Finally,the popularity information is forgotten using orthogonal loss,and the next recommendation task is performed.However,this method is model fixed and cannot be applied to other models.Therefore,the first method in this article cannot solve the depolarization needs of most base models.Aiming at the problem of the first method,this article proposes a second depolarization method.This article uses the classic Skip-Gram embedding algorithm’s popularity negative sampling method.This negative sampling method was originally designed to accelerate computation,but it was found in this article that it can effectively alleviate popularity bias.This article uses this negative sampling method to construct a model independent depolarization framework,which can ensure the accuracy of the base model under this framework,Achieving good results in popularity indicators,addressing the shortcomings of our first model.This article also proposes two indicators to measure the prevalence bias of models.Effective comparative experiments have been conducted on multiple datasets and base models,which can prove that the model proposed in this article can effectively alleviate the prevalence bias and provide effective solutions for the academic and industry.