Fairness in Supervised Learning: An Information Theoretic Approach

Abstract

Automated decision making systems are increasingly being used in real-world applications. In these systems for the most part, the decision rules are derived by minimizing the training error on the available historical data. Therefore, if there is a bias related to a sensitive attribute such as gender, race, religion, etc. in the data, say, due to cultural/historical discriminatory practices against a certain demographic, the system could continue discrimination in decisions by including the said bias in its decision rule. We present an information theoretic framework for designing fair predictors from data, which aim to prevent discrimination against a specified sensitive attribute in a supervised learning setting. We use equalized odds as the criterion for discrimination, which demands that the prediction should be independent of the protected attribute conditioned on the actual label. To ensure fairness and generalization simultaneously, we compress the data to an auxiliary variable, which is used for the prediction task. This auxiliary variable is chosen such that it is decontaminated from the discriminatory attribute in the sense of equalized odds. The final predictor is obtained by applying a Bayesian decision rule to the auxiliary variable.

Can Information Flows Suggest Targets for Interventions in Neural Circuits?

“…For ANNs specifically, these can also take the form of visualizations to describe what features or what abstract concepts an ANN has learned [23,24]. There have also been a number of information-theoretic approaches for measuring bias and promoting fairness in AI systems [25][26][27][28][29].…”

“…It is also straightforward to extend our work to alternative definitions of bias such as Equalized Odds[28,50] by suitably modifying Definition 2. However, it is not clear if such a measure can still be meaningfully interpreted as an "information flow".…”

An Information-Theoretic Quantification of Discrimination with Exempt Features

“…In a sense, this work treads a middle ground between two schools of thought, namely, demographic parity (Agarwal et al 2018;Ghassami, Khodadadian, and Kiyavash 2018), which enforces the criterion Z ⊥ ⊥Ŷ , and equalized odds (Hardt et al 2016;Ghassami, Khodadadian, and Kiyavash 2018), which enforces Z ⊥ ⊥Ŷ |Y (directly or through practical relaxations) where Y denotes the true labels of the historic dataset. Our selective quantification of non-exempt discrimination helps address one of the major criticisms against demographic parity, namely, that it can deliberately choose unqualified members from the protected group (Zemel et al 2013), e.g., by disregarding the critical features if they are correlated with Z.…”

Sociolectal Analysis of Pretrained Language Models

“…Alternatively, we can think of the language model with the lowest divergence across groups as the most fair language model; such a definition differs from standard Rawlsian fairness, but has been advocated inKamishima et al (2012);Ghassami et al (2018).…”