Work

Addresing interpretability fairness & privacy in machine learning through combinatorial optimization methods

Year: 2023
Type: dissertation
Abstract: Machine learning techniques are increasingly used for high-stakes decision making, such as college admissions, loan attribution or recidivism prediction. It is thus crucial to ensure that the trained models can be audited or understood by human users, do not create or reproduce discrimination or bias, and do not leak sensitive information regarding their training data. Indeed, interpretability, fairness and privacy are key requirements for trustworthy machine learning, and all three have been studied extensively during the last decade. However, they are often considered in isolation. The objective of this manuscript is precisely to investigate the interactions between the three different fields, using tools from combinatorial optimization and operations research. For each pairwise intersection, we review the literature on the previously observed tensions or synergies, and draw new insights from our proposed contributions - either highlighting an identified tension or proposing a conciliation mechanism. We first propose an Integer Linear Programming (ILP) based pruning technique for a learning algorithm producing fair and inherently interpretable models. By leveraging jointly accuracy, sparsity and fairness, it enhances the exploration of the algorithm's search space and helps conciliating fairness and interpretability. Motivated by the empirical observation that fairness often does not generalize well, we further introduce a novel framework to improve fairness robustness with respect to the training set sampling. We then show how information regarding a model's fairness can be exploited to reconstruct its training set sensitive attributes. To this end, we propose efficient MILP and Constraint Programming (CP) models directly encoding the fairness information to improve the reconstruction performed by any baseline adversary. This highlights the intrinsic tension between enforcing fairness with respect to some sensitive attributes, and ensuring such attributes' privacy. Finally, we demonstrate how the structure of a released interpretable model can be used to reconstruct a probabilistic version of its training set. By precisely quantifying the amount of information a model encodes regarding its training data, we illustrate an apparent conflict between interpretability and privacy considerations.
Source: HAL
Language: English
Open Access status: green
OpenAlex ID: W7045346913
Cites: 233
Cited by: 2
Related to: 8