Fairness-Aware Processing Techniques in Survival Analysis: Promoting Equitable Predictions

Zhouting Zhao, Tin Lok James Ng

Abstract: As machine learning (ML) systems are becoming pervasive in high-stakes applications, the issue of ML fairness is receiving increasing attention. A large variety of fair ML solutions have been developed to ensure that bias and inaccuracies in the data and model do not lead to decisions that treat individuals unfavorably on the basis of sensitive characteristics. While most of the fair ML literature focus on classification and regression setting, fairness of survival analysis for time-to-event outcomes are under-explored. In contrast to existing fair survival analysis solutions which typically incorporate fairness constraints in the learning mechanisms, we propose several pre-processing and post-processing approaches. Due to the model-agnostic nature of pre-processing and post-processing methods, they may offer more flexible fairness intervention. Additionally, pre-processing and post-processing methods tend to be more intuitive and explainable compared to in-processing methods. We carry out experimental studies with medical and non-medical data sets to evaluate the proposed fairness methods.