One-vs.-One Mitigation of Intersectional Bias: A General Method to Extend Fairness-Aware Binary Classification

TL;DR

The paper tackles intersectional bias in binary classification by introducing One-vs.-One Mitigation (OVOM), a general framework that performs pairwise subgroup comparisons and aggregates their mitigations to produce per-instance scores and subgroup-specific thresholds. OVOM is designed to work with pre-, in-, and post-processing fairness methods and to optimize accuracy under a user-defined disparity cap, \\epsilon, while tracking metrics such as \\gamma_d and \\gamma_r for multiple fairness criteria. Empirical evaluation on the Adult and COMPAS datasets across demographic parity, equalized odds, and equal opportunity shows OVOM consistently reduces subgroup disparities relative to conventional methods, with varying effects on accuracy depending on dataset and method. Overall, OVOM broadens the applicability of fairness-aware binary classification to settings with multiple sensitive attributes and provides a controllable trade-off between fairness and accuracy, facilitating more realistic and equitable decision-making.

Abstract

With the widespread adoption of machine learning in the real world, the impact of the discriminatory bias has attracted attention. In recent years, various methods to mitigate the bias have been proposed. However, most of them have not considered intersectional bias, which brings unfair situations where people belonging to specific subgroups of a protected group are treated worse when multiple sensitive attributes are taken into consideration. To mitigate this bias, in this paper, we propose a method called One-vs.-One Mitigation by applying a process of comparison between each pair of subgroups related to sensitive attributes to the fairness-aware machine learning for binary classification. We compare our method and the conventional fairness-aware binary classification methods in comprehensive settings using three approaches (pre-processing, in-processing, and post-processing), six metrics (the ratio and difference of demographic parity, equalized odds, and equal opportunity), and two real-world datasets (Adult and COMPAS). As a result, our method mitigates the intersectional bias much better than conventional methods in all the settings. With the result, we open up the potential of fairness-aware binary classification for solving more realistic problems occurring when there are multiple sensitive attributes.

Figures (4)

• Figure 1: A toy example of intersectional bias. Grey circles are the applicants for loan applications, and those surrounded by black squares are the accepted ones. The percentages are the acceptance rate in each group. In this example, the disparate impact (DI), which is the ratio of the acceptance rate of a protected group to that of a non-protected group is used as the fairness metrics. In US law, it is said that if the value of disparate impact is more than 0.8, there is not an unfair situation (80% rule) 10.1145/2783258.2783311. However, in this example, Even if the fairness metrics are satisfied in each sensitive attribute (DI is more than 0.8 for both gender and race), there is a subgroup that is clearly discriminated, non-white female, whose acceptance rate is 0%.
• Figure 2: The overview diagram of our One-vs.-One Mitigation. Our method uses a bias mitigation function corresponding to each subgroup. When there are four subgroups $S=\{I,I\space I, I\space I\space I, I\space V\}$, there are six subgroup pairs. In this example, we compare the results based on three subgroup pairs (I-I I, I-I I I, and I-I V) because $Xi$ belongs to I. Our method aggregates the mitigation results on each pairs, and calculates the $T_i$ based on a voting rate of the favorite class and the average value of the predicted probabilities. The final mitigation result $Zi$ is decided by whether the voting rate exceeds a threshold value $\theta(s)$.
• Figure 3: Comparison of balanced accuracy and disparity for each method. (a)The disparity of $\gamma_d$ on Adult. (b)The disparity of ratio $\gamma_r$ on Adult. (c)The disparity of difference $\gamma_d$ on COMPAS. (d)The disparity of ratio $\gamma_r$ on COMPAS. the X- and Y-axes represent disparity and balanced accuracy respectively. Since the ideal value of the disparity is 0, and that of the balanced accuracy is 1, the further to the upper left a point is positioned, the better the result. In each figure, the left column represents demographic parity, the middle column represents equalized odds, and the right column represents equal opportunity. The upper row represents pre-processing, the middle row represents in-processing, and the bottom row represents post-processing. All points and error bars represent the mean and standard deviations respectively, for 5-fold cross-validation.
• Figure 8: Trade-off between balanced accuracy and $\gamma$ with difference values of $\epsilon$ on Adult dataset. The X-axis, left Y-axis, and right Y-axis represent $\epsilon$, balanced accuracy, and disparity, respectively. The dotted line represents disparity, which is measured with the right Y-axis, corresponding to the value of $\epsilon$, and, in the desired result, the disparity is lower than the line. Top row: Mitigation results for demographic parity by AD with our method. Bottom row: Mitigation results for equalized odds by EO with our method. The dotted line represents $\epsilon$, and it is desirable that the disparity is less than that line.

Theorems & Definitions (5)

• Definition 3.1: Concept of subgroup fairness criteria
• Definition 3.2: Demographic Parity
• Definition 3.3: Equalized Odds
• Definition 3.4: Equal Opportunity
• Definition 5.1: Subgroup Disparity