Title: Low-rank approximations with fairness constraints
Authors: Emanuele Aliverti - University of Padova (Italy) [presenting]
David Dunson - Duke University (United States)
Abstract: In many high-dimensional applications, there is considerable interest in developing machine learning algorithms that are designed to achieve some notion of fairness. These considerations are crucial when algorithms aid or replace human judgement; for example, in criminal justice risk assessment or medical imaging. When standard methods fail in characterizing with efficiency and flexibility the explosion in the number of dimensions, common strategies for reducing the number of parameters include sparse modelling, latent structure analysis and low-rank factorization; however, without explicit adjustment, there are no guarantees that such methods will also be fair. The focus is on including fairness constraints in low-dimensional representation of the original data, in order to preserve fairness guarantees with respect to sensitive attributes such as race or sex, while maintaining accuracy of the approximation. Efficient computational algorithms and theoretical support for the approaches will be discussed, along with empirical results in a variety of application.