FairerML: An Extensible Platform for Analysing, Visualising, and Mitigating Biases in Machine Learning

作     者：Yuan, Bo Gui, Shenhao Zhang, Qingquan Wang, Ziqi Wen, Junyi Mao, Bifei Liu, Jialin Yao, Xin 

作者机构：Southern Univ Sci & Technol Shenzhen Peoples R China Huawei Technol Co Ltd Shenzhen Peoples R China 

出 版 物：《IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE》 (IEEE Comput. Intell. Mag.)

年 卷 期：2024年第19卷第2期

页      面：129-141页

核心收录：

学科分类：0808[工学-电气工程] 08[工学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：National Key R#x0026 D Program of China 

主　　题：Training Measurement Analytical models Machine learning algorithms Computational modeling Data visualization Machine learning 

摘      要：Given the growing concerns about bias in machine learning, dozens of metrics have been proposed to measure the fairness of machine learning. Several platforms have also been developed to compute and illustrate fairness metrics on platform-provided data. However, most platforms do not provide a user-friendly interface for users to upload and evaluate their own data or machine learning models. Moreover, no known platform is capable of training machine learning models, while considering their fairness and accuracy simultaneously. Motivated by the above insufficiency, this work develops FairerML, an extensible platform for analysing, visualising, and mitigating biases in machine learning. Three core functionalities are implemented in FairerML: fairness analysis of user-uploaded datasets, fairness analysis of user-uploaded machine learning models, and the training of a set of Pareto models considering accuracy and fairness metrics simultaneously by using multiobjective learning. The clear visualisation and description of the fairness analysis and the configurable model training process of FairerML make it easy for training fairer machine learning models and for educational purposes. In addition, new fairness metrics and training algorithms can be easily integrated into FairerML thanks to its extendability.