Abstract
Effective machine learning models can automatically learn useful information from a large quantity of data and provide decisions with high accuracy. These models may, however, lead to unfair predictions, which can be harmful for specific demographic groups. This paper introduces threshold invariant fairness, a method that ensures a model remains equitable even if the decision cutoff is shifted, achieved by aligning the underlying risk distributions of different demographic groups.
Introduction
In high-stakes machine learning—such as predicting criminal recidivism (COMPAS) or loan defaults—fairness is often measured using "classification parity." Common metrics include Demographic Parity (ensuring equal selection rates across groups) and Equalized Odds (ensuring equal error rates).
However, these metrics are typically calculated at a default decision threshold (e.g., 0.5). If a decision-maker shifts that threshold to be more conservative or lenient, a model that was "fair" at 0.5 can suddenly become significantly biased. True fairness should be a property of the model's scoring distribution, not just a single point on a curve.
Methodology
The authors propose that to achieve threshold invariance, we must equalize the risk distributions (the probability density of scores) across protected groups. If the curves representing different groups have the same shape, any threshold slice will yield a fair result. They introduce two fairness regularizers:
- Histogram Approximation (HA): Since standard histograms are not differentiable, the authors use Gaussian kernels to estimate the density of scores in various "bins." They then minimize the Kullback–Leibler (KL) divergence between the histograms of different groups.
- Gaussian Assumption (GA): A simpler approach that assumes the raw scores follow a normal distribution. The model is trained to match the means and variances of these distributions across groups.
These regularizers were integrated into standard classifiers, including Logistic Regression, Linear SVM, and Kernel SVM.
Experimental Setup
The study primarily utilizes the COMPAS risk assessment dataset to predict whether a defendant will recommit a crime within two years. The protected attribute is race (African-American vs. Caucasian). Features include age, gender, priors count, and charge degree. The data was split into 70% training and 30% testing sets, with continuous features normalized and categorical features 0-1 encoded.
Key Results and Discussion
Experiments demonstrated that the proposed methods provide significantly more stability across thresholds compared to traditional point-alignment methods. While baseline models showed bias fluctuations as thresholds moved, the HA and GA models maintained consistently low bias.
Visualizations show that under "Strong Constraints," the risk distribution curves for African-American and Caucasian groups almost perfectly overlap, effectively neutralizing the impact of the decision threshold. Interestingly, the simpler Gaussian Assumption (GA) performed nearly as well as the more complex HA while being computationally more efficient.
Limitations and Future Work
Achieving threshold invariance involves an accuracy-fairness tradeoff; fair models saw a drop in accuracy compared to the baseline. Additionally, the current work focuses on binary protected attributes and assumes Gaussian distributions for the GA method, which may not hold for all datasets. Future research will explore adapting these methods for deep learning and multi-group scenarios.
Summary updated