Bayesian Models and Algorithms for Fairness and Transparency

No photo
BCAM principal investigator: Novi Quadrianto
BCAM research line(s) involved:
Reference: 851538 - BayesianGDPR (H2020)
Coordinator: U.SUSSEX - University of Sussex and BCAM - Basque Center for Applied Mathematics
Partners: Host Institution: U.SUSSEX - University of Sussex
Beneficiaries: U.SUSSEX - University of Sussex
BCAM - Basque Center for Applied Mathematics
Duration: 2021 - 2025
Funding agency: ERCEA - Starting Grants 2019
Type: International Project
Status: Ongoing Project


EU's GDPR prescribes that ""Personal Data shall be processed lawfully, fairly, and in a transparent manner."" The vision of this BayesianGDPR project is to integrate into automated machine learning systems using a novel Bayesian approach, in a transparent manner, the legal non-discriminatory principles of GDPR, taking into account feedback from humans and future consequences of their outputs. We aim to achieve this ambitious vision by 1) developing a machine learning framework for addressing fairness in classification problems and beyond, and under uncertainty about data, models, and predictions about future data (algorithmic fairness under uncertainty), 2) extending the framework to a setting where data points arrive over time, and models have to be dynamically updated when taking general feedback (feedback-driven setting), and 3) ensuring a human could understand how non-discrimination is defined and achieved by using, among others, uncertainty estimates for building interpretable models and/or explicitly explaining about changes being made to the models to enforce non-discriminatory principles (transparency in fairness). The BayesianGDPR project is ""doubly timely""; not just in terms of the criticality of the fairness and transparency in machine learning at this point in time, but also because recent breakthroughs in scalability have finally made it feasible to explore Bayesian approaches that are uniquely capable of addressing one of the most central aspects of the problem, i.e. uncertainty. BayesianGDPR will, in the short term, ensure that organisations relying on machine learning technologies are provided with concrete tools to comply with the non-discriminatory principles of GDPR and similar laws. In the medium term, it will impact research in computational law, and its integration into mainstream legal practice. In the long term, it will also ensure continued confidence of the general public in the deployment of machine learning systems.

Related Projects