Towards Fair Unsupervised Learning

Francois Buet-Golfouse, Islam Utyagulov

Abstract: Bias-mitigating techniques are now well established in the supervised learning literature and have shown their ability to tackle fairness-accuracy, as well as fairness-fairness trade-offs. These are usually predicated on different conceptions of fairness, such as demographic parity or equal odds that depend on the available labels in the dataset. However, it is often the case in practice that unsupervised learning is used as part of a machine learning pipeline (for instance, to perform dimensionality reduction or representation learning via SVD) or as a standalone model (for example, to derive a customer segmentation via k-means). It is thus crucial to develop approaches towards fair unsupervised learning. This work investigates fair unsupervised learning within the broad framework of generalised low-rank models (GLRM). Importantly, we introduce the concept of fairness functional that encompasses both traditional unsupervised learning techniques and min-max algorithms (whereby one minimises the maximum group loss). To do so, we design straightforward alternate convex search or biconvex gradient descent algorithms that also provide partial debiasing techniques. Finally, we show on benchmark datasets that our fair generalised low-rank models (“fGLRM”) perform well and help reduce disparity amongst groups while only incurring small runtime overheads.