去除推荐场景多混淆因子的因果去偏方法  

Causal Debias Method for Multiple Confounding Factors in Recommendation

作　　者：杨庚杭 沈苏彬[2] YANG Geng-hang;SHEN Su-bin(School of Computer Science,Nanjing University of Posts and Telecommunications,Nanjing 210023,China;National Engineering Research Center on Communication and Networking,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)

机构地区：[1]南京邮电大学计算机学院,江苏南京210023 [2]南京邮电大学通信与网络技术国家工程研究中心,江苏南京210003

出　　处：《计算机技术与发展》2024年第9期9-15,共7页Computer Technology and Development

基　　金：国家自然科学基金(62002174)。

摘　　要：如何去除推荐场景中存在的偏差问题,是推荐系统提升效果的重大挑战。现有的推荐模型只是拟合数据,基于相关性去除偏差会受到虚假相关性的影响;基于因果关系去除偏差则由于场景复杂很难抽取更加全面的因果关系。因此,尽可能多地考虑各因素之间的因果关系并去除偏差问题很有必要。该文从因果角度对用户行为在物品分类的分布不平衡和物品类别流行度两个混淆因子在推荐流程中的因果关系进行研究,提出考虑多因素的去混淆方法,有效去除推荐中的偏差问题。首先,分析各变量之间的因果关系并构建因果图;其次,使用前门调整和后门调整去除两个混淆因子造成的虚假相关性以及偏差;最后,将该方法应用在神经网络因子分解机上,在两个公开数据集上进行了实验并验证。从仿真实验结果可知,该方法相比于目前的最优方法都有不同程度的提升。How to eliminate the bias problem existing in recommendation is a major challenge for recommender systems to improve their effectiveness.The existing recommendation models only fit the data,and eliminating bias based on correlation may be affected by spurious correlations.Eliminating bias based on causal relationships is difficult to extract more comprehensive causal relationships due to the complex scene.Therefore,it is necessary to consider the causal relationships between various factors as much as possible and eliminating bias issues.We study the causal relationship of the two confounding factors in the recommendation process,namely,the distribution imbalance of user behavior in item classification and item category popularity.A de-confound method is proposed that considers multiple factors to effectively eliminate the bias problem in recommendation.Firstly,the causal relationship between variables in the recommendation is analyzed and a causal graph is constructed.Secondly,the false correlation as well as the bias caused by the two confounding factors are eliminated using front-door adjustment and back-door adjustment.Finally,the proposed method is applied to neural factorization machines,and experimentally validated on two publicly available datasets,the results show that the proposed method has varying degrees of improvement compared to the current optimal method.

关 键 词：因果推断 前门调整 后门调整 多混淆因子 因果图 

分 类 号：TP391.3[自动化与计算机技术—计算机应用技术]