On the relation between accuracy and fairness in binary classification

By Indrė Žliobaitė. 2015

206 citations recorded by the source corpus.

Our study revisits the problem of accuracy-fairness tradeoff in binary classification. We argue that comparison of non-discriminatory classifiers needs to account for different rates of positive predictions, otherwise conclusions about performance may be misleading, because accuracy and discrimination of naive baselines on the same dataset vary with different rates of positive predictions. We provide methodological recommendations for sound comparison of non-discriminatory classifiers, and present a brief theoretical and empirical analysis of tradeoffs between accuracy and non-discrimination.

Subfields: Artificial Intelligence, AI Ethics, Fairness & Society

How researchers see this paper

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

In Artificial Intelligence

In AI Ethics, Fairness & Society

Citations in the corpus

Fair Prediction with Disparate Impact: A Study of Bias in Recidivism Prediction Instruments (2016, neutral): “That is, if ( 2 : 5 ) Statistical parity also goes by the name of equal acceptance rates 14 or group fairness , 15 although it should be noted that these terms are in many cases not used synonymously.”
On Fairness and Calibration (2017, background): “It has been shown that, under most frameworks of fairness, there is a trade-off between algorithmic performance and non-discrimination [4, 9, 19, 39].”
Weakly-Supervised Disentanglement Without Compromises (2020, background): “While they observed a strong correlation between demographic parity (Calders et al., 2009; Zliobaite, 2015) and disentanglement, the applicability of their approach is limited by the fact that disentangled representations are difficult to identify without access to explicit observations of the…”
On the Fairness of Disentangled Representations (2019, background): “To evaluate the learned representations r(x) of these observations, we assume that the set of ground-truth factors of variation include both a target factor y, which we would like to predict from the learned representation, and an underlying sensitive factor s, which we want to be fair to in the sen”
Learning Fair Rule Lists (2019, methodology): “For instance in this paper, we have used statistical parity [10, 21, 23, 33, 66] as fairness notion.”
Learning Controllable Fair Representations (2018, background): “For conciseness, we focus on the demographic parity notion of fairness (Calders et al., 2009; Zliobaite, 2015; Zafar et al., 2015), which requires the decisions made by a classifier over z to be independent of the sensitive attributes u.”
Fairness in Machine Learning: A Survey (2020, unknown): “Consequently, users must decide on where to place emphasis, but be mindful of the trade off between any fairness measure and model accuracy [26, 87, 73, 129, 296, 54].”
Fairness Definitions Explained (2018, unknown): “Groupfairness[12](a.k.a. statisticalparity[12],equal acceptance rate [24], benchmarking [9]) .”
Measuring discrimination in algorithmic decision making (2017, background): “The normalized difference (Zliobaite 2015) is the mean difference for binary classiﬁ-cation normalized by the rate of positive outcomes, where ) .”
Convex Formulations for Fair Principal Component Analysis (2018, background): “…have studied how to deﬁne fairness for supervised learning (Hardt, Price, and Srebro 2016; Calders, Kamiran, and Pechenizkiy 2009; Dwork et al. 2012; Zliobaite 2015) and developed supervised learners that maintain high prediction accuracy while reducing unfairness (Berk et al. 2017; Chouldechova…”
Benchmarking Bias Mitigation Algorithms in Representation Learning through Fairness Metrics (2021, unknown): “Calders and Verwer ([52]) 1− [p(Ŷ = 1|S = 1)− p(Ŷ = 1|S 6= 1)] CV Demographic parity ([49] & [27]) = Group fairness ([17]) = statistical parity ([17]) = equal acceptance rate ([53]) p(Ŷ ) = p(Ŷ |S) equivalent to 1−CV equivalent to p(Ŷ = 1|S = 1) = p(Ŷ = 1|S 6= 1) DP”
Spectral Algorithms for Computing Fair Support Vector Machines (2017, background): “This definition is typically stated [Calders et al., 2009, Zliobaite, 2015, Zafar et al., 2017] in terms of a single threshold, though it can be generalized to multiple thresholds.”