文章目录
1. 偏差分析
2. 数据偏差
2.1. 选择偏差 Selection Bias
2.2. 一致性偏差 Conformity Bias
2.3. 曝光偏差 Exposure Bias
2.4. 位置偏差 Position Bias
3. 模型偏差
3.1. 归纳偏差 Inductive Bias
4. 推荐结果的偏差与不公平性
4.1. 流行度偏差 Popularity Bias
4.2. 偏见 Unfairness (用户偏差 User Bias)
正文
1. 偏差分析
(1) Bias-Variance Decomposition for Ranking. WSDM 2021;
(2) Transfer Learning in Collaborative Recommendation for Bias Reduction. RecSys 2021;
code: https://csse.szu.edu.cn/staff/panwk/publications/TJR/
2. 数据偏差
2.1. 选择偏差 Selection Bias
用户的显式反馈往往很稀疏，并且只有点击的样本才可能有显式反馈。假如用户对样本的反馈是打分，当用户点击样本时，意味着用户已经比较喜欢该样本了，因此打分可能偏高，若不合预期，打分也可能偏低，此时用户的反馈存在选择偏差。这部分观测数据打分的分布，和全体样本的打分分布是有较大差异的，相当于采样不随机，不能很好预估总体。
(1) Measuring and Mitigating Item Under-Recommendation Bias in Personalized Ranking Systems. SIGIR 2020;
(2) E-commerce Recommendation with Weighted Expected Utility. CIKM 2020;
code: https://github.com/zhichaoxu-shufe/E-commerce-Rec-with-WEU.
(3) Combating Selection Biases in Recommender Systems with a Few Unbiased Ratings. WSDM 2021;(4) Non-Clicks Mean Irrelevant Propensity Ratio Scoring As a Correction. WSDM 2021;
(5) Mitigating Confounding Bias in Recommendation via Information Bottleneck. RecSys 2021;
code: https://github.com/dgliu/RecSys21_DIB .
(6) Pessimistic Reward Models for Off-Policy Learning in Recommendation. RecSys 2021.
2.2. 一致性偏差 Conformity Bias
人是社会性动物，在推荐系统中也是如此。同样以打分为例，用户的评分很容易受大众/朋友的影响，大家都说好的电影，可能你觉得没那么好，但由于不想那么突出，因此也选择从众。这就导致样本数据无法反映用户真实的偏好。
(1) Debiasing Item-to-Item Recommendations With Small Annotated Datasets. RecSys 2020;
code: https://github.com/microsoft/debiasing-item2item .
2.3. 曝光偏差 Exposure Bias
相对于全量索引，推荐系统返回的结果是极其有限的，用户的反馈也只是在这部分曝光样本上产生，对于更多未曝光的样本，由于缺少反馈数据，导致模型无法很好学习未曝光数据的分布。模型训练时是在曝光样本空间，但是预估时是全量样本空间，这就导致曝光偏差。曝光偏差对于新物品不太友好，模型容易对旧物品打高分。
(1) Unbiased Learning for the Causal Effect of Recommendation. RecSys 2020;
code: https://www.dunnhumby.com/source-files/.
(2) Debiased Explainable Pairwise Ranking from Implicit Feedback. RecSys 2021;
code: https://github.com/KhalilDMK/EBPR .
(3) Top-K Contextual Bandits with Equity of Exposure. RecSys 2021;
code: https://github.com/deezer/carousel_bandits .
2.4. 位置偏差 Position Bias
推荐物品的位置也会影响用户的点击率。出于对平台推荐内容的信任，以及用户的使用习惯，位置靠前的视频，往往更容易被点击，但并不代表用户真的喜欢。
(1) Attribute-based Propensity for Unbiased Learning in Recommender Systems Algorithm and Case Studies. KDD 2020;
(2) Unbiased Ad Click Prediction for Position-aware Advertising Systems. RecSys 2020;
(3) Unbiased Learning to Rank in Feeds Recommendation. WSDM 2021;
code: https://github.com/flamewei123/Unbaised-LTR-in-Feeds-Recommendation-WSDM21 .
(4) Cross-Positional Attention for Debiasing Clicks. WWW 2021;
3. 模型偏差
3.1. 归纳偏差 Inductive Bias
归纳偏差来源于模型，定义模型时为了简化问题会人为增加假设，在预测时可能产生泛化误差。
(1) A General Knowledge Distillation Framework for Counterfactual Recommendation via Uniform Data. SIGIR 2020;
code: https://github.com/dgliu/SIGIR20_KDCRec .
(2) Counterfactual Evaluation of Slate Recommendations with Sequential Reward Interactions. KDD 2020;
code: https://github.com/spotify-research/RIPS_KDD2020 .
4. 推荐结果的偏差与不公平性
4.1. 流行度偏差 Popularity Bias
即长尾效应。通常推荐系统分发内容中，头部曝光现象严重，导致样本分布不均匀。热门样本数量更多，这也导致模型倾向给热门样本打高分，分发频率远超其他。通常会对热门样本进行降采样，缓解长尾效应。
1) Keeping Dataset Biases out of the Simulation : A Debiased Simulator for Reinforcement Learning based Recommender Systems. RecSys 2020;
code: https://github.com/BetsyHJ/SOFA .
(2) Popularity-Opportunity Bias in Collaborative Filtering. WSDM 2021;
(3) Diverse User Preference Elicitation with Multi-Armed Bandits. WSDM 2021;
(4) ProtoCF: Prototypical Collaborative Filtering for Few-shot Item Recommendation. RecSys 2021;
code: https://github.com/aravindsankar28/ProtoCF .
(5) Analyzing Item Popularity Bias of Music Recommender Systems: Are Different Genders Equally Affected? RecSys 2021;
(6) The Idiosyncratic Effects of Adversarial Training on Bias in Personalized Recommendation Learning. RecSys 2021;
code: https://github.com/sisinflab/The-Idiosyncratic-Effects-of-Adversarial-Training .
(7) Biases in Recommendation System. RecSys 2021;
4.2. 偏见 Unfairness (用户偏差 User Bias)
人工智能的伦理问题近年来热度也比较高，例如模型在人种、性别、年龄方面存在歧视，本质上是样本不均匀导致。推荐系统的偏见来源于数据，数据不够多样不够充分，导致模型无法学习充分，在预估时产生了偏见。
(1) Debiasing Career Recommendations with Neural Fair Collaborative Filtering. WWW 2021;
code: https://github.com/rashid-islam/nfcf .
(2) User Bias in Beyond-Accuracy Measurement of Recommendation Algorithms. RecSys 2021;
(3) Measuring and Mitigating Bias and Harm in Personalized Advertising. RecSys 2021;
(4) I Want to Break Free! Recommending Friends from Outside the Echo Chamber. RecSys 2021;
(5) Leave No User Behind Towards Improving the Utility of Recommender Systems for Non-mainstream Users. WSDM 2021;
code: https://github.com/roger-zhe-li/wsdm21-mainstream .
