Wasserstein Fair Classification

By Ray Jiang, Aldo Pacchiano, T. Stepleton, Heinrich Jiang, Silvia Chiappa. 2019

216 citations recorded by the source corpus.

We propose an approach to fair classification that enforces independence between the classifier outputs and sensitive information by minimizing Wasserstein-1 distances. The approach has desirable theoretical properties and is robust to specific choices of the threshold used to obtain class predictions from model outputs. We introduce different methods that enable hiding sensitive information at test time or have a simple and fast implementation. We show empirical performance against different fairness baselines on several benchmark fairness datasets.

Subfields: Safety, Robustness & Interpretability

How researchers see this paper

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

In Safety, Robustness & Interpretability

Citations in the corpus

Reducing Sentiment Bias in Language Models via Counterfactual Evaluation (2019, methodology): “This consideration leads us to match output distributions, which is called “Strong Demographic Parity” (Jiang et al., 2019).”
Fairness in Machine Learning: A Survey (2020, background): “When extending the classifier’s (convex) loss function with fairness terms, researchers typically seek to balance fairness and accuracy (e.g., References [5, 26, 36, 52, 57, 98, 119, 141, 153, 168, 208, 314]).”
Fairness in Machine Learning (2020, methodology): “To deal with these limitations, the work in [27, 84] introduced the Strong Demographic Parity (SDP) criterion, which requires 𝑆 | = 𝐴 , i.e.”
A General Approach to Fairness with Optimal Transport (2020, methodology): “In Jiang et al. (2019), we introduced an approach to fair classiﬁcation that uses optimal transport theory to match the distributions of the model outputs corresponding to different sensitive attributes to a common distribution.”
Learning to Score Behaviors for Guided Policy Optimization (2019, methodology): “Distance Metrics: WDs have been used in many applications in machine learning where guarantees based on distributional similarity are required [13, 1].”
Your fairness may vary: Pretrained language model fairness in toxic text classification (2021, methodology): “…between the group-wise true positive rates) (Chzhen et al., 2019), using optimal transport to ensure similarity in group-wise predicted score distributions (Jiang et al., 2020), and a plug-in approach for transforming the predicted probabilities to satisfy fairness constraints (Yang et al., 2020).”
Mitigating Political Bias in Language Models Through Reinforced Calibration (2021, methodology): “Now, the indirect bias is computed using the distance be-tween the base rates of X o and X ∀ o ∈ A : where ∆ BR is the second order Sliced Wasserstein Distance (SWD) (Jiang et al. 2019; Rabin et al. 2011) between the base rates (computed by f judge ) of two sets of texts.”
Benchmarking Bias Mitigation Algorithms in Representation Learning through Fairness Metrics (2021, methodology): “In-processing techniques [6, 7, 8, 9, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33] try to remove sensitive information from the learned representation space.”
Quantifying Social Biases in NLP: A Generalization and Empirical Comparison of Extrinsic Fairness Metrics (2021, background): “(2020a) choose Wasserstein-1 distance (Jiang et al., 2020) 3 21 , while Borkan et al.”