基于自监督学习的去流行度偏差推荐方法  

Self-Supervised Learning for Alleviating Popularity Bias in Recommender Systems

作　　者：张帅[1,2] 高旻[1,2] 文俊浩[1,2] 熊庆宇[1,2] 唐旭[2] ZHANG Shuai;GAO Min;WEN Jun-hao;XIONG Qing-yu;TANG Xu(Key Laboratory of Dependable Service Computing in Cyber Physical Society(Chongqing University),Ministry of Education,Chongqing 400044,China;School of Big Data&Software Engineering,Chongqing University,Chongqing 400044,China)

机构地区：[1]信息物理社会可信服务计算教育部重点实验室(重庆大学),重庆400044 [2]重庆大学大数据与软件学院,重庆400044

出　　处：《电子学报》2022年第10期2361-2371,共11页Acta Electronica Sinica

基　　金：国家自然科学基金(No.72161005);重庆市自然科学基金(No.cstc2020jcyj-msxmX0690);重庆大学中央高校基本科研业务费项目(No.2020CDJ-LHZZ-039);重庆市技术创新与应用发展专项重点项目(No.cstc2019jscx-fxydX0012);重庆市留学人员创业创新支持计划(No.cx2020097)。

摘　　要：近年来,随着推荐系统研究的不断深入,推荐系统的公平性受到越来越多关注.流行度偏差也即流行的物品比非流行的物品更容易被推荐,是影响其公平性的重要因素之一.流行度偏差对推荐系统的各利益相关者都有严重的影响,引起研究者的广泛关注.相关研究主要通过推荐结果重排或学习过程中融合正则化项提升非流行物品的曝光率,而非流行物品的交互数据极度稀疏成为研究的瓶颈.针对此问题,本文提出基于自监督学习的去流行度偏差推荐方法,解决两个难点:(1)准确学习交互数据极度稀疏的非流行物品的表征;(2)提升非流行物品曝光率的同时,兼顾不同用户对流行和非流行物品的偏好.具体地,从用户的角度,提出流行物品和非流行物品双视图的用户偏好学习方法,准确学习用户对流行和非流行物品的真实偏好;从物品的角度,采用自监督学习,利用互信息最大化捕获非流行物品与流行物品间的潜在关系,辅助提升非流行物品嵌入学习的准确性.最后,设计用户流行度偏好一致性、资格公平性等指标,并通过三个公开数据集的大量实验说明了本文方法在提升推荐性能的同时,能有效缓解流行度偏差问题并具有较强的通用性.In recent years,with the development of recommender system,more and more attention has been paid to the fairness of recommender system.Popularity bias mens that popular items are more likely to be recommended than un⁃popular items,which is one of the important factors affecting its fairness.Popularity bias has a serious impact on all stake⁃holders in the recommendation system,which has aroused widespread concern of researchers.Related research mainly im⁃proves the exposure rate of unpopular items by the rearrangement of recommended results or the integration of regulariza⁃tion items in the learning process,but the extremely sparse interaction data of unpopular items becomes the bottleneck of re⁃search.To solve this problem,this paper proposes self-supervised learning for alleviating popularity bias(SSLAB)to deal with two difficulties:(1)accurately learning the representation of unpopular items with extremely sparse interactive data;(2)improving the exposure rate of unpopular items while taking into account the preferences of different users for popular and unpopular items.Specifically,from the perspective of users,this paper proposes a dual view user preference learning method for popular and unpopular items to accurately learn users’real preferences for popular and unpopular items;from the perspective of items,self-supervised learning is used to capture the potential relationship between unpopular items and popular items by maximizing mutual information to help improve the accuracy of unpopular items embedded learning.Fi⁃nally,metrics such as user popularity preference consistency and qualification fairness are designed,and a large number of experiments on three open data sets show that the proposed method can effectively alleviate the popularity bias and has a strong scalability while improving the recommendation performance.

关 键 词：推荐系统 协同过滤 公平性 流行度偏差 自监督学习 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]