A1513
Title: Blockwise-fairness gap penalty for subgroup-fair clustering
Authors: Jeong Hankyo - Seoul National University (Korea, South) [presenting]
Abstract: AI fairness is a critical concern in automated decision-making systems. A fair clustering algorithm is proposed for settings with multiple sensitive attributes. Subgroup-blocks are introduced to represent sensitive groups at multiple scales, ranging from marginal groups to subgroups, and a fairness metric called the blockwise-fairness gap is defined over a given collection of subgroup-blocks. Because this gap is discrete and non-smooth, it is difficult to optimize directly. Therefore a smooth surrogate fairness gap and an efficient algorithm that minimizes clustering cost while penalizing the surrogate fairness gap is developed, yielding simple closed-form updates and enabling numerically stable optimization. Theoretical guarantees are provided showing that reducing the proposed surrogate fairness gap reduces the blockwise-fairness gap. Experiments on benchmark datasets demonstrate that the proposed method achieves favorable cost-fairness trade-offs, with computational efficiency and numerical stability compared to existing approaches.
