基于因果推断的推荐系统去偏研究综述

A Survey on Debiasing Recommendation Based on Causal Inference

推荐系统是解决信息过载的重要技术之一.然而,推荐系统中存在各种各样的偏差问题,影响了对用户真实偏好的建模,制约了推荐性能的提升.近年来,因果推断理论的发展为分析和解决推荐系统偏差问题提供了有力的支持.因果推断是一种从观测数据中识别变量之间因果关系和估计因果效应的统计学方法,通过构建和分析因果模型,帮助推荐系统识别和消除偏差,提高对用户偏好拟合的准确性.本文对基于因果推断的推荐系统去偏研究的主要工作进行了全面的综述.本文根据推荐系统的不同阶段将偏差的产生分为三个阶段;首先,概述了因果推断的原理和方法,并阐述了因果推断与推荐去偏之间的联系,为缓解偏差问题提供了思路;接着,针对每一阶段的偏差,探讨了现有的因果推断技术如何应用于推荐去偏,分类和归纳了现有的因果推荐去偏方法,并进行了详细的对比分析;最后,讨论和展望了基于因果推断的推荐系统去偏研究未来的发展趋势.

Recommender systems play a vital role in addressing information overload by learning user preferences from historical interaction data and thereby providing personalized recommendations.However,various biases in recommendation systems hinder the accurate modeling of users' true preferences,limiting the improvement of recommendation performance.Recently,the development of causal inference theory has provided robust support for analyzing and resolving bias problems in recommender systems.Causal inference,a statistical method aimed at identifying and estimating causal effects between variables from observational data,assists in identifying and eliminating biases through the construction and analysis of causal models,enhancing the accuracy of fitting user preferences.Applying causal inference to debiasing tasks in recommender systems has achieved significant success,effectively mitigating bias while also enhancing accuracy and reliability.This paper provides a comprehensive review of the research on debiasing recommendations based on causal inference.Considering that bias occurs at various stages of recommender systems,we classify the sources of bias according to the three stages of recommender systems:data,algorithms,and evaluation.We also summarize the manifestations of bias at each stage and their impact on recommendations.Based on the study of debiasing recommendations from a causal perspective,we first outline the principles and key methods of causal inference.This establishes the connection between causal inference and debiasing recommendation,providing insights into mitigating bias.Then we systematically organize and analyze debiasing strategies for recommender systems at the data,algorithm,and evaluation stages based on causal inference techniques.For debiasing methods at the data stage,there are primarily two strategies based on how the data is utilized:counterfactual construction-based methods generate synthetic data points to simulate what might happen under different scenarios,helping to uncover hidden biases;unbiased data-based methods involve collecting data that is free from the common biases present in real-world datasets.For debiasing methods at the algorithm stage,there are primarily three strategies based on different causal techniques:causal representation learning-based methods aim to learn representations of the data that are invariant to biases;causal intervention-based methods directly manipulate variables to observe changes and infer causal relationships;counterfactual reasoning-based methods involve comparing actual outcomes with hypothetical scenarios to identify and correct biases.For debiasing methods at the evaluation stage,there are primarily two strategies based on the correction and optimization of unbiased estimates:inverse propensity scoring-based methods adjust for the probability of receiving a particular treatment,helping to balance the dataset;doubly robust-based methods combine propensity score weighting with outcome modeling to improve the robustness and accuracy of bias correction.Currently,recommender systems based on causal inference represent a novel and challenging research field.This paper summarizes several open research directions,including debiasing recommendation methods based on causal discovery,a general causal-based debiasing recommendation framework,robust debiasing methods based on causal inference,addressing bias issues in dynamic environments using causal approaches,and the construction of datasets for causal debiasing recommendations.Finally,we summarize this paper and provide an outlook on the research of debiasing recommendations based on causal inference from the perspectives of application needs and technological development.