Fair Collaborative Learning (FairCL): A Method to Improve Fairness amid Personalization

Model personalization has attracted widespread attention in recent years. In an ideal situation, if individuals’ data are sufficient, model personalization can be realized by building models separately for different individuals using their own data. But, in reality, individuals often have data sets of varying sizes and qualities. To overcome this disparity, collaborative learning has emerged as a generic strategy for model personalization, but there is no mechanism to ensure fairness in this framework. In this paper, we develop fair collaborative learning (FairCL) that could potentially integrate a variety of fairness concepts. We further focus on two specific fairness metrics, the bounded individual loss and individual fairness, and develop a self-adaptive algorithm for FairCL and conduct both simulated and real-world case studies. Our study reveals that model fairness and accuracy could be improved simultaneously in the context of model personalization. History: Bianca Maria Colosimo served as the senior editor for this article. Funding: This work was supported by the Breakthrough T1D Award [Grant 2-SRA-2022-1259-S-B]. Data Ethics & Reproducibility Note: The code capsule is available on Code Ocean at https://codeocean.com/capsule/1331847/tree/v1 and in the e-Companion to this article (available at https://doi.org/10.1287/ijds.2024.0029 ). The real-world data, including the transportation demand management and surgical site infection data sets, are proprietary and not publicly available. Other results are available at https://github.com/ryanlif/FairCL .