Abstract
How can we control for latent discrimination in predictive models? How can we provably remove it? Such questions are at the heart of algorithmic fairness and its impacts on society. In this paper, we define a new operational fairness criteria, inspired by the well-understood notion of omitted variable-bias in statistics and econometrics. Our notion of fairness effectively controls for sensitive features and provides diagnostics for deviations from fair decision making. We then establish analytical and algorithmic results about the existence of a fair classifier in the context of supervised learning. Our results readily imply a simple, but rather counter-intuitive, strategy for eliminating latent discrimination. In order to prevent other features proxying for sensitive features, we need to include sensitive features in the training phase, but exclude them in the test/evaluation phase while controlling for their effects. We evaluate the performance of our algorithm on several real-world datasets and show how fairness for these datasets can be improved with a very small loss in accuracy.
Introduction
The standard approach to algorithmic fairness is "fairness through blindness"—simply removing sensitive attributes (protected features) from the dataset. However, this often fails due to latent discrimination. In many datasets, non-sensitive features (like zip code or education level) are highly correlated with protected attributes (like race or income). If you remove the protected attribute, a machine learning model will naturally use the correlated features as "proxies" to reconstruct that missing information and minimize loss. This is known in statistics as Omitted Variable Bias (OVB).
This paper introduces the Train-Then-Mask algorithm, which moves beyond vague descriptions to a formal, operational definition called EL-fairness (Explicit and Latent fairness) to eliminate these proxy effects.
The Train-Then-Mask Strategy
The authors propose a counter-intuitive strategy: to stop a model from using proxies for race or gender, you must first show the model exactly what race and gender look like. If the model can "see" the sensitive attribute during training, it won't try to "guess" it using other variables.
- Training: Train the "optimal unconstrained classifier" ($h^*$) using all available features, including the sensitive ones. By including the sensitive feature, the model assigns the "discriminatory" signal directly to that feature, preventing other features from being forced to act as proxies.
- Masking: During the testing or deployment phase, the sensitive feature is "masked" by setting it to a fixed constant (e.g., setting gender to '0' for everyone).
- Thresholding: Because masking the feature shifts the model's output, the authors introduce a parameter $ au$ (a threshold shift) and find the value $ au^*$ that maximizes accuracy on a validation set.
Key Insights and Results
The theoretical guarantees of this method rely on the classifier being "separable," meaning the influence of the sensitive feature can be mathematically isolated. For a broad class of separable classifiers (including Linear SVMs and Logistic Regression), this two-step process produces the optimal fair classifier.
Experimental Performance
The researchers evaluated their approach on three standard benchmarks: the Adult Income, German Credit, and COMPAS Recidivism datasets.
- Superior Fairness: In experiments using Linear SVMs, the Train-Then-Mask approach achieved zero latent discrimination ($L_{Discr} = 0.000$) across all three datasets, significantly outperforming "omitting the feature" or "data massaging" techniques.
- Minimal Accuracy Trade-off: On the Adult Income dataset, the unconstrained model had 82.5% accuracy. The Train-Then-Mask model achieved 82.3% accuracy—a negligible loss for a significant gain in fairness.
Conclusion and Future Work
Train-Then-Mask provides a robust framework for eliminating latent discrimination. While the formal mathematical proof of optimality holds for separable functions, empirical tests on non-separable Neural Networks showed that the algorithm still significantly reduced latent discrimination compared to standard baselines. Future work will explore measuring the "degree of non-separability" in complex models and more sophisticated masking techniques to further neutralize sensitive features during inference.
Summary updated