Improving Model Fairness for a More Trustworthy AI
Maintaining fairness in AI models is a pressing concern in the evolving field of artificial intelligence (AI). Two research areas stand out in this context: how to ensure AI fairness in federated learning (a method for training AI models with decentralized data) and the impact of data distribution shifts during the training and deployment of AI models. Professor Kangwook Lee’s team at the University of Wisconsin-Madison has been investigating these challenging issues in AI fairness and seeking ways to improve AI fairness despite data distribution shifts.
In a recent work presented at IEEE ISIT 2023 [1], the team evaluated two popular federated fair learning schemes: Local Fair Training (LFT) combined with an ensemble method and LFT combined with Federated Averaging (LFT+FedAvg). The findings indicate that LFT+FedAvg offers better fairness than the ensemble method, albeit at the cost of more frequent communication. However, LFT+FedAvg may not achieve the level of optimal fairness possible through centralized training.
This discovery helps explain the success of some recently proposed federated fair learning algorithms that use additional communication rounds. The team also provided numerical experiments conducted in various general settings to support their findings.
In another work presented at ICML 2023 [2], Prof. Lee and his collaborators at KAIST addressed the limitations of in-processing fair algorithms. Their research highlighted a common but flawed assumption in many fairness techniques that training and deployment data distributions are identical. The team demonstrated analytically how this assumption, coupled with shifts in the bias between labels and sensitive groups, can negatively impact a model’s fairness.
To address these challenges, the team proposed a novel pre-processing step to sample input data in a way that reduces correlation shifts, thereby improving the performance of in-processing techniques. They introduced an optimization problem that adjusts the data ratio among labels and sensitive groups to mirror the shifted correlation, effectively improving the effectiveness of in-processing fair algorithms. Tested on synthetic and real datasets, this novel approach consistently improved accuracy and fairness.
Prof. Lee’s work, along with his collaborators, marks a significant advance in machine learning fairness. Their research offers valuable insights and practical solutions that can guide the development of more fair and accurate AI systems.
References
[1] Yuchen Zeng, Hongxu Chen, and Kangwook Lee, “Federated Learning with Local Fairness Constraints”, ISIT 2023
[2] Yuji Roh, Kangwook Lee, Steven Euijong Whang, Changho Suh, “Improving Fair Training under Correlation Shifts”, ICML 2023
Reported by Dr. Kangwook Lee (2022 KSEA YIG Grant Recipient)
KSEA e-Letters 