Class-Wise Group DRO for Fairness

• The paper introduces class-wise group DRO, performing adversarial reweighting separately for each label to enforce fairness via Equalized Conditional Accuracy.
• It establishes a connection between Difference of Conditional Accuracy and variance of groupwise losses, enabling a principled fairness regularization framework.
• Empirical results on tabular, vision, and language tasks show that adaptive reweighting with smoothed updates improves both convergence and the accuracy-fairness trade-off.

Searching arXiv for the focal paper and closely related group DRO work to ground the article with current citations. arxiv_search(query="(Jung et al., 2023) Re-weighting Based Group Fairness Regularization via Classwise Robust Optimization", max_results=5) arxiv_search(query="class-wise group distributionally robust optimization fairness Equalized Conditional Accuracy DCA", max_results=10) Class-wise group distributionally robust optimization is a class-conditional variant of group distributionally robust optimization in which the adversarial reweighting over sensitive groups is performed separately within each label class rather than over pooled groups. In the formulation introduced in "Re-weighting Based Group Fairness Regularization via Classwise Robust Optimization" (Jung et al., 2023), the method is designed for supervised classification with inputs $x \in \mathcal{X}$, class labels $y \in \mathcal{Y}$, and sensitive group labels $a \in \mathcal{A}$, and it serves as a fairness-aware training framework that unifies two lines of work: re-weighting underrepresented groups and fairness regularization based on explicit group-fairness metrics. Its central claim is that, when fairness is defined conditionally on the true class, class-wise DRO gives a principled objective whose adversarial inner problem produces adaptive group weights while the outer problem minimizes a balanced empirical risk plus a penalty tied to Difference of Conditional Accuracy (DCA) (Jung et al., 2023).

1. Problem setting and conceptual role

The basic setting uses a labeled dataset

$\mathcal{D}=\{(x_i,y_i,a_i)\}_{i=1}^N,$

together with group- and class-restricted subsets $\mathcal{D}_a$ and $\mathcal{D}_a^y$. The empirical loss is defined as

$$\mathcal{L}(\theta,\mathcal{D}) \triangleq \frac{1}{|\mathcal{D}|}\sum_{i=1}^{|\mathcal{D}|}\ell(\theta,(x_i,y_i)).$$

The distinctive feature of the class-wise formulation is that it does not treat group robustness as a single worst-group problem over all samples. Instead, it decomposes the objective by class and performs robust reweighting within each set $\{\mathcal{D}_a^y\}_{a\in\mathcal{A}}$. In the fairness paper, this decomposition is motivated by the fact that the target notion of fairness is conditional on the true class, so a pooled group-DRO formulation does not preserve the same semantics (Jung et al., 2023).

This places class-wise group DRO between standard empirical risk minimization and classical worst-group optimization. ERM minimizes the average per-instance loss and can hide systematic disparities across groups. Standard group DRO minimizes the worst group loss over a flat grouping. Class-wise group DRO retains adversarial group weighting, but restricts the competition among groups to be class-conditional. A plausible implication is that the method is best suited to settings where the relevant disparity criterion is explicitly conditioned on the label rather than only on marginal group performance.

2. Equalized Conditional Accuracy and Difference of Conditional Accuracy

The fairness notion used in the class-wise formulation is Equalized Conditional Accuracy (ECA): for every class $y$, all groups should have equal conditional prediction accuracy (Jung et al., 2023). The violation of this condition is measured by Difference of Conditional Accuracy (DCA),

$$\Delta_{\text{DCA}} \triangleq \frac{1}{|\mathcal{Y}|}\sum_{y\in\mathcal{Y}}\Delta_y, \qquad \Delta_y \triangleq \max_{a,a'}\left|P(\widehat{Y}=y\mid A=a,Y=y)-P(\widehat{Y}=y\mid A=a',Y=y)\right|.$$

For binary classification, ECA matches Equalized Odds, and in the multi-class case it generalizes Equal Opportunity-like parity across groups for each class (Jung et al., 2023). This class-conditional interpretation is the reason the method is not merely a generic robust optimization scheme: the fairness target itself is indexed by $y \in \mathcal{Y}$0.

A common misconception is to equate any group-aware reweighting with a fairness regularizer. In the focal formulation, the fairness interpretation depends on the specific correspondence between DCA and the dispersion of class-conditional group losses. Without that correspondence, worst-group reweighting remains a robustness device, but it no longer has the same direct status as a surrogate for ECA.

3. Variance characterization and the class-wise DRO objective

A central technical result in the fairness formulation is an equivalence, under the $y \in \mathcal{Y}$1-$y \in \mathcal{Y}$2 loss $y \in \mathcal{Y}$3, between empirical DCA and the variance of groupwise class-conditional losses. For each class $y \in \mathcal{Y}$4,

$y \in \mathcal{Y}$5

The same variance admits a pairwise-loss-gap representation,

$y \in \mathcal{Y}$6

The DRO side starts from the standard group DRO objective

$y \in \mathcal{Y}$7

and then replaces the simplex-only adversary by a $y \in \mathcal{Y}$8-divergence uncertainty set that permits quasi-probabilities,

$y \in \mathcal{Y}$9

For this uncertainty set,

$a \in \mathcal{A}$0

Applying this DRO construction separately for each class yields the class-wise group DRO objective

$a \in \mathcal{A}$1

Combining the DCA-variance equivalence with the Group DRO-variance equivalence gives the regularized form

$a \in \mathcal{A}$2

where $a \in \mathcal{A}$3 (Jung et al., 2023).

This is the formal basis for viewing the method as fairness regularization rather than only robustness to group shift. The class-wise decomposition is therefore not a cosmetic change to standard group DRO; it is the mechanism that aligns the min-max objective with a class-conditional fairness criterion.

4. Adaptive reweighting, quasi-probabilities, and optimization

The inner maximization yields adaptive group weights $a \in \mathcal{A}$4 for each class, with a closed form for the optimum $a \in \mathcal{A}$5 (Jung et al., 2023). The paper notes that these weights may be negative, which is allowed by the quasi-probability uncertainty set. This is important because negative weights let the method penalize over-performing groups more aggressively than standard positive re-weighting schemes.

The resulting interpretation is dual. As a DRO regularizer, the objective imposes pressure against disparities in class-conditional group losses. As a re-weighting method, it generates learned group weights rather than heuristic reweighting coefficients. This unifies the two fairness paradigms emphasized in the paper: explicit metric-aware regularization and adaptive reweighting (Jung et al., 2023).

Optimization is carried out by an iterative alternating procedure. At epoch $a \in \mathcal{A}$6, model parameters are updated by gradient descent on the weighted classwise loss,

$a \in \mathcal{A}$7

The adversarial weights are then updated for each class by solving the inner maximization,

$a \in \mathcal{A}$8

Instead of exponentiated-gradient ascent, which would require simplex constraints, the method uses a smoothed Iterated Best Response update because $a \in \mathcal{A}$9 permits quasi-probabilities. In practice, $\mathcal{D}=\{(x_i,y_i,a_i)\}_{i=1}^N,$0 is trained with mini-batch cross-entropy, while $\mathcal{D}=\{(x_i,y_i,a_i)\}_{i=1}^N,$1 is updated using full-dataset group losses and smoothed over epochs with $\mathcal{D}=\{(x_i,y_i,a_i)\}_{i=1}^N,$2. The paper states that this smoothing stabilizes otherwise oscillatory weight dynamics and improves convergence and final DCA (Jung et al., 2023).

5. Empirical behavior and ablation findings

The reported experiments cover tabular, vision, and language tasks: Adult, COMPAS, UTKFace, CelebA, and CivilComments (Jung et al., 2023). On these benchmarks, the class-wise DRO method is reported to consistently give the best accuracy-fairness trade-off, often achieving lower DCA than strong baselines while preserving competitive accuracy.

The ablation results isolate several ingredients. Directly optimizing the fairness surrogate without DRO is worse, and using DRO without class-wise treatment is also suboptimal. Allowing quasi-probabilities further improves fairness, and smoothing the Iterated Best Response updates improves convergence and final DCA (Jung et al., 2023). The method is also described as scalable and adaptable because it can be inserted into standard training pipelines, uses ordinary gradient descent for model updates, and avoids expensive full retraining or constrained saddle-point solvers.

The empirical emphasis on class-wise decomposition is consistent with broader observations about group DRO. In binary tabular imbalance learning, group DRO with class labels used as groups outperformed ERM and several classical imbalance methods on average $\mathcal{D}=\{(x_i,y_i,a_i)\}_{i=1}^N,$3-mean and ROC-AUC across 19 datasets, especially when imbalance is severe and classes overlap substantially (Mustapha et al., 2023). That case study is not a fairness paper, but it supports the broader point that worst-group objectives over class-defined groups can materially change optimization behavior relative to ERM.

At the same time, the focal fairness paper makes a stricter claim: collapsing all $\mathcal{D}=\{(x_i,y_i,a_i)\}_{i=1}^N,$4 pairs into a single group DRO problem weakens the correspondence to DCA and empirically worsens performance (Jung et al., 2023). In that sense, class-wise group DRO is not simply group DRO with finer groups; it is tied to the class-conditional structure of the target fairness notion.

6. Relation to the broader group DRO literature

Class-wise group DRO belongs to a wider family of methods that alter how groups are represented, how adversarial weights are chosen, or where robustness is enforced.

Approach	Group structure	Salient property
Class-wise DRO fairness regularization (Jung et al., 2023)	Hard groups within each class $\mathcal{D}=\{(x_i,y_i,a_i)\}_{i=1}^N,$5	Links DRO to DCA and ECA
PG-DRO (Ghosal et al., 2023)	Soft class-environment groups	Replaces one-hot group labels with probabilistic membership
Common-good weighting (Piratla et al., 2021)	Group losses with simplex weights	Chooses weights by gradient agreement rather than worst-group emphasis
Generalized / function-space group DRO (Soma et al., 2022, Thudi et al., 2024)	Group distributions with uncertainty set $\mathcal{D}=\{(x_i,y_i,a_i)\}_{i=1}^N,$6 or convex hull mixtures	Extends worst-group DRO to top-$\mathcal{D}=\{(x_i,y_i,a_i)\}_{i=1}^N,$7, CVaR, weighted ranking, and optimal mixtures

Several nearby variants clarify what is specific to the class-wise fairness setting. PG-DRO uses probabilistic group membership rather than hard annotations; in that formulation, groups are class-environment pairs $\mathcal{D}=\{(x_i,y_i,a_i)\}_{i=1}^N,$8, and the method reduces exactly to standard G-DRO when the group probabilities become one-hot (Ghosal et al., 2023). This makes probabilistic groups a soft version of class-wise group DRO. By contrast, "Focus on the Common Good: Group Distributional Robustness Follows" (Piratla et al., 2021) argues that classical worst-group emphasis can overfocus on idiosyncratic hard groups and instead chooses group weights based on aggregate gradient agreement, prioritizing updates that help many groups at once. That paper is group-aware but not class-wise in the DCA sense.

Other work broadens the optimization lens. "Near-Optimal Algorithms for Group Distributionally Robust Optimization and Beyond" (Soma et al., 2022) formulates a generalized min-max problem

$\mathcal{D}=\{(x_i,y_i,a_i)\}_{i=1}^N,$9

where $\mathcal{D}_a$0 can encode worst-group risk, empirical CVaR, average top-$\mathcal{D}_a$1 group loss, or weighted ranking over sorted group losses. "MixMax: Distributional Robustness in Function Space via Optimal Data Mixtures" (Thudi et al., 2024) reparameterizes group DRO in function space and shows that, for bounded functions and losses such as cross-entropy and mean squared error, the robust predictor can be obtained through an optimal mixture distribution in the convex hull of the groups. These developments suggest that class-wise group DRO is one member of a broader design space in which the adversary may operate over hard groups, soft groups, mixtures of groups, or ranked subsets of groups.

A distinct line of critique concerns evaluation rather than training. "Re-evaluating Group Robustness via Adaptive Class-Specific Scaling" (Seo et al., 2024) shows that robust and average accuracies lie on a clear Pareto trade-off and that a training-free class-specific scaling of logits can make a naive ERM baseline match or even outperform recent debiasing methods on robust metrics. This does not refute class-wise DRO fairness regularization, because the focal fairness method optimizes DCA rather than only worst-group or unbiased accuracy, but it underscores that group-robust methods should be evaluated across the full robustness-accuracy landscape rather than at a single operating point.

Finally, when explicit group labels are unavailable, "Modeling the Q-Diversity in a Min-max Play Game for Robust Optimization" (Wu et al., 2023) learns latent group assignments and augments minority groups through mixing, retaining the worst-group objective while replacing annotated groups with pseudo-groups. A plausible implication is that future class-wise DRO systems may combine the class-conditional fairness structure of (Jung et al., 2023) with soft or learned group definitions in the style of PG-DRO and Q-Diversity.