基于帕累托效应视角下的推荐系统多角度公平性  

A Multi-Side Fairness-Aware Recommendation System Based on a Pareto-Efficient Perspective

作　　者：杜清月 黄晓雯 桑基韬 DU Qingyue;HUANG Xiaowen;SANG Jitao(School of Computer and Information Technology, Beijing Key Lab of Traffic Data Analysis and Mining, Beijing Jiaotong University, Beijing 100044, China)

机构地区：[1]北京交通大学计算机与信息技术学院,交通数据分析与挖掘北京市重点实验室,北京100044

出　　处：《太原理工大学学报》2022年第1期89-97,共9页Journal of Taiyuan University of Technology

基　　金：中央高校基本科研专项基金资助项目(2021RC217)。

摘　　要：提出了一种从帕累托视角解决推荐系统多角度公平性的方法,通过对抗正则化器消除用户嵌入中的敏感属性信息,采用基于曝光的负采样策略提高推荐系统的准确率,从而达到帕累托最优。并且,基于曝光的负采样策略在一定程度上解决物品曝光偏差的问题,保证了物品角度公平性,实现了用户、物品的多角度公平性。实验结果表明,该方法在保证推荐准确率的同时有效提高了用户和物品角度的公平性。In this paper proposed was a method to solve the multi-side fairness of the recommender system from the perspective of Pareto.It eliminates the sensitive attribute information in the user embedding through the adversarial regularizer,and adopts the negative sampling strategy based on exposure to improve the accuracy of the recommender system,so as to achieve Pareto optimality.In addition,the exposure-based negative sampling strategy solves the problem of item exposure bias to a certain extent,ensures the fairness of item side,and realizes the multi-side fairness of users and items.Experimental results show that the method effectively improves the fairness of users and items while ensuring the accuracy of recommendation.

关 键 词：推荐系统公平性 多角度公平性 帕累托效应 对抗学习 负采样 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]