Conceitos essenciais
Achieving group fairness in downstream models through pre-processing with FairRR.
Resumo
This article discusses the importance of fairness in machine learning models and introduces FairRR, a pre-processing algorithm that aims to achieve group fairness by modifying response variables. The paper highlights the connection between group fairness metrics and optimal design matrices, emphasizing the control of disparity levels while maintaining model utility.
INTRODUCTION
- Increased use of machine learning in decision-making processes raises concerns about algorithmic fairness.
- Various approaches have been developed to ensure fairness, including pre-processing, in-processing, and post-processing methods.
PRE-PROCESSING FOR GROUP FAIRNESS
- Proposes FairRR as a pre-processing algorithm to modify response variables for achieving group fairness.
- Discusses the theoretical foundation connecting group fairness metrics with optimal design matrices.
RANDOMIZED RESPONSE AND FAIRNESS CONSTRAINTS
- Introduces Randomized Response as a privacy technique to modify labels based on probabilities.
- Shows how measures of group fairness can be controlled through flipping response variables based on sensitive attributes.
EXPERIMENTS AND RESULTS
- Evaluates FairRR's performance on benchmark datasets for fair classification.
- Compares FairRR with other pre-processing methods and demonstrates its ability to control disparity levels effectively.
CONCLUSION AND FUTURE RESEARCH
- Concludes that FairRR is an efficient and theory-motivated algorithm for achieving group fairness in machine learning models.
- Suggests future research directions include generalizing FairRR to multiple sensitive attributes and exploring its relationship with privacy mechanisms.
Estatísticas
Fair Sampling Kamiran and Calders [2012]は、すべての4つのグループのサイズを調整する方法に基づいています。
TabFairGAN RajabiとGaribay[2021]は、一般的なWGANモデルのジェネレータ損失にフェアネスペナルティ項を追加します。
FAWOS Salazar et al.[2021]は、敏感属性の分布に基づくフェアネス対応オーバーサンプリングアルゴリズムです。