Proxy Non-Discrimination in Data-Driven Systems

By Anupam Datta, Matt Fredrikson, Gihyuk Ko, Piotr (Peter) Mardziel, Shayak Sen. 2017

59 citations recorded by the source corpus.

Machine learnt systems inherit biases against protected classes, historically disparaged groups, from training data. Usually, these biases are not explicit, they rely on subtle correlations discovered by training algorithms, and are therefore difficult to detect. We formalize proxy discrimination in data-driven systems, a class of properties indicative of bias, as the presence of protected class correlates that have causal influence on the system's output. We evaluate an implementation on a corpus of social datasets, demonstrating how to validate systems against these properties and to repair violations where they occur.

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

Citations in the corpus

FairSight: Visual Analytics for Fairness in Decision Making (2019, background): “In addition, a proxy variable is a variable correlated with a sensitive attribute whose use in a decision procedure can result in indirect discrimination [9].”
Does mitigating ML's disparate impact require disparate treatment? (2017, background): “At what point is the disparate treatment simply correcting for biased labels? At what point does it more explicitly amount to affirmative action? Recent work on identifying proxy discrimination [DFK17] and causal formulations of fairness [NS17, KRCP17, KLRS17] offer approaches to framing such proble”
Does mitigating ML's impact disparity require treatment disparity? (2017, background): “Recent work on identifying proxy discrimination [31] and causal formulations of fairness [32–34] offer some promising approaches to translating such understanding into technological solutions.”
The explanation game: a formal framework for interpretable machine learning (2020, unknown): “For instance, a model may not explicitly include a protected attribute such as race, but instead use a seemingly innocuous covariate like zip code, which is often a strong predictor of race (Datta et al. 2017).”
The Disparate Effects of Strategic Manipulation (2018, methodology): “As such, in this paper we ask, “What are the effects of strategic classification and manipulation in a world of social stratification?” As in previous work in strategic classification, we cast the problem as a Stackelberg game in which the learner moves first and publishes her classifier before cand”
Fairness-Aware Programming (2019, background): “[6, 7], where they consider influence of a proxy.”
The Measure and Mismeasure of Fairness: A Critical Review of Fair Machine Learning (2018, background): “Recently, researchers have further expanded the idea of anti-classiﬁcation to create algorithms that avoid potentially suspect causal paths (Bonchi et al., 2017; Datta et al., 2017; Kilbertus et al., 2017; Kusner et al., 2017; Nabi and Shpitser, 2018).”
Bias in machine learning - what is it good for? (2020, background): “One problem with this approach is that the result may still be biased with respect to race, if other features are strongly correlated with race and therefor act as proxies for race in the learning [14, 45].”