Controllable Universal Fair Representation Learning

Learning fair and transferable representations of users that can be used for a wide spectrum of downstream tasks (specifically, machine learning models) has great potential in fairness-aware Web services. Existing studies focus on debiasing w.r.t. a small scale of (one or a handful of) fixed pre-defined sensitive attributes. However, in real practice, downstream data users can be interested in various protected groups and these are usually not known as prior. This requires the learned representations to be fair w.r.t. all possible sensitive attributes. We name this task universal fair representation learning, in which an exponential number of sensitive attributes need to be dealt with, bringing the challenges of unreasonable computational cost and un-guaranteed fairness constraints. To address these problems, we propose a controllable universal fair representation learning (CUFRL) method. An effective bound is first derived via the lens of mutual information to guarantee parity of the universal set of sensitive attributes while maintaining the accuracy of downstream tasks. We also theoretically establish that the number of sensitive attributes that need to be processed can be reduced from exponential to linear. Experiments on two public real-world datasets demonstrate CUFRL can achieve significantly better accuracy-fairness trade-off compared with baseline approaches.