Fair Representations by Compression

Organizations that collect and sell data face increasing scrutiny for the discriminatory use of data. We propose a novel unsupervised approach to map into compressed binary representation independent sensitive attributes. show in an information bottleneck framework, parsimonious should filter out related attributes if they are provided directly decoder. Empirical results method achieves state-of-the-art accuracy-fairness trade-off explicit control entropy bit stream allows user move smoothly simultaneously along both rate-distortion rate-fairness curves.