FeDa4Fair
收藏资源简介:
FeDa4Fair是一个用于生成表格数据集的库,专门用于在异构客户端偏差下评估公平的联邦学习(FL)方法。该库的目标是解决FL中公平性的挑战,其中客户端之间存在着不同的数据偏差,导致模型对不同客户端的公平性不同。FeDa4Fair支持在客户端级别进行公平性评估,以解决现有FL方法在单一敏感属性上的局限性。FeDa4Fair基于美国社区调查的公共使用微观数据样本(ACS PUMS)和Ding等人提出的公平相关收入和就业预测任务(ACSIncome和ACSEmployment)。该数据集涵盖了美国51个州的数据,其中人口和社会经济状况各不相同,从而模拟了真实的异构客户端设置。FeDa4Fair提供了数据集不可知的方法来评估和放大现有偏差,以实现广泛的偏差异构客户端场景。FeDa4Fair支持单个和交集公平性评估,允许根据单个属性或属性组合进行偏差评估。FeDa4Fair目前支持两种公平性度量标准,即人口差异(DD)和均等化机会差异(EOD),用于进行评估。
FeDa4Fair is a library for generating tabular datasets, specifically designed to evaluate fair federated learning (FL) methods under heterogeneous client bias. The library aims to address the fairness challenges in FL, where disparate data biases across clients lead to inconsistent model fairness across different client groups. FeDa4Fair supports client-level fairness evaluation to address the limitations of existing FL methods that only focus on single sensitive attributes. Built upon the Public Use Microdata Samples (PUMS) of the American Community Survey (ACS) and the fairness-relevant income and employment prediction tasks (ACSIncome and ACSEmployment) proposed by Ding et al., FeDa4Fair covers data from all 51 U.S. states with varying demographic and socioeconomic characteristics, thus simulating real-world heterogeneous client settings. FeDa4Fair provides dataset-agnostic methods to evaluate and amplify existing biases, enabling the exploration of a wide range of biased heterogeneous client scenarios. It supports both single and intersectional fairness evaluations, allowing bias assessment based on individual attributes or combinations of multiple attributes. Currently, FeDa4Fair supports two fairness metrics: Demographic Disparity (DD) and Equalized Odds Difference (EOD) for evaluation purposes.




