Group Fairness Constraints in ML

• Group Fairness Constraints are formal mathematical conditions that ensure equitable, statistically balanced outcomes across defined demographic groups.
• They are implemented using methods like constrained optimization, penalty-based approaches, spectral techniques, and discrete algorithms to maintain fairness in various ML applications.
• Empirical studies demonstrate that these constraints can achieve near-parity across groups with minimal utility loss, even in large-scale, complex environments.

Group Fairness Constraints refer to formal restrictions or requirements imposed within machine learning, optimization, or resource allocation models to ensure that outcomes across defined demographic, sensitive, or policy-relevant groups are equitable according to a specified fairness metric. These constraints manifest as explicit mathematical conditions, typically relating to representation, distributional equity, or parity of outcomes. The rigorous design and enforcement of group fairness constraints underpins fairness-centric algorithmic research across supervised learning, combinatorial optimization, reinforcement learning, clustering, matching, and more.

1. Formal Definitions and Taxonomy

Group fairness constraints are stated with respect to a partitioning of the instance or decision space into $k$ or more groups $\{G_1,\ldots, G_k\}$ along protected attributes (e.g., race, gender, age). The constraint can be encoded in different modalities, summarized as follows:

Domain	Constraint Type	Formal Example / Metric
Classification (ML)	Statistical Parity	$$|\mathbb{P}[h(x)=1|x\in G_i] - \mathbb{P}[h(x)=1|x\in G_j]| \leq \varepsilon$$ (Zhang et al., 2021)
Clustering	Proportional Group Representation	$B_h|C_i| \leq |C_{i,h}| \leq a_h|C_i|$ for all clusters $i$, all groups $h$ (Dickerson et al., 2023)
Regression	Group Means Equality	$$|\mathbb{E}[f(x)|x\in G_i] - \mathbb{E}[f(x)|x\in G_j]| = 0$$ (Fitzsimons et al., 2018, Fitzsimons et al., 2018)
Bipartite Matching	Integral Quotas or Upper/Lower Bounds	$$\ell_{j,k} \leq \sum_{i\in I_k} x_{i,j} \leq U_{j,k}$$ (per class and platform) (Sankar et al., 2021, Panda et al., 2022)
Knapsack/Resource Allocation	Value/Weight/Cardinality Bounds	$L^v_g \leq \sum_{i\in G_g} v_i x_i \leq U^v_g$ (per group $g$) (Patel et al., 2020)
Influence Maximization	Minimum Fractional Spread	$\min_i I_{G,C_i}(A)/|C_i| \geq \tau$ for all groups $C_i$ (Tsang et al., 2019)
RL, Policy Optimization	Max Gap in Return Across Groups	$|J(\pi_i; \cdot) - J(\pi_j; \cdot)| \leq \epsilon$ for all $i, j$ (Song et al., 10 Mar 2025, Deng et al., 2022)
Optimal Transport	Blockwise Mass Constraints	$\pi_{SW}(s, w) = F_{sw}$ for all group pairs (Bleistein et al., 12 Jan 2026)

Statistical parity and proportionality are canonical, but group fairness also encompasses constraints on group-wise accuracy, exposure (ranking), resource allocation, and complex multi-group intersectionality.

2. Mathematical Formulations and Motivation

The mathematical encoding of group fairness constraints is domain-specific but adheres to a set of unifying principles:

• Distributional Parity: Requiring the distribution of outcomes (predictions, assignments, rewards) to match across all groups, either exactly or within a specified slack. This is exemplified by statistical parity and group mean constraints (Zhang et al., 2021, Fitzsimons et al., 2018).
• Quota-Based Constraints: Upper and/or lower bounds on group representation within clusters, matchings, or selections; e.g., each demographic group $g$ must be assigned at least $\ell_g$ and at most $u_g$ units of the resource (Dickerson et al., 2023, Patel et al., 2020, Sankar et al., 2021).
• Proportionality/Balance: Enforcing group proportions in clusters or selections to mirror those in the full dataset, within tolerance $\delta$ (Dickerson et al., 2023, Ojeda-Ruiz et al., 22 Oct 2025, Wang et al., 2022).
• Fairness Metrics in Embedding Space: Defining group fairness in terms of disparities in distance or spread (e.g., via Gini coefficient across group-wise embeddings in GNNs) (Sirohi et al., 2024).
• Cross-Group Coupling/Marginals: In matching/OT, exact mass is assigned between group pairs as specified by a target matrix, as in $\pi_{SW}(s,w) = F_{sw}$ for all $(s, w)$ (Bleistein et al., 12 Jan 2026).

Motivations for enforcing group fairness constraints include legal mandates (disparate impact), social welfare considerations (Rawlsian, Nash social welfare), multi-group intersectionality, and the empirical prevention of under-representation or discrimination of minority populations.

3. Algorithmic Approaches to Enforcing Group Fairness

Algorithmic instantiations and enforcement strategies vary significantly by domain, but common methodologies include:

1. Constrained Optimization Most settings cast group fairness constraints as explicit constraints in an optimization program, such as integer programs, convex programs, or constrained policy optimization (Patel et al., 2020, Song et al., 10 Mar 2025, Deng et al., 2022, Zhang et al., 2021).
2. Penalty-Based/Lagrangian Methods Constraints are moved into the objective via Lagrange multipliers, possibly regularizing for constraint violation (Sirohi et al., 2024, Bleistein et al., 12 Jan 2026). Automatic balancing and adaptive tuning of penalty weights is used to modulate the fairness–utility trade-off.
3. Augmented Spectral and Linear Algebraic Frameworks In clustering/spectral decomposition, fairness is encoded as a set of linear or quadratic constraints incorporated into trace-minimization eigenproblems. Techniques for avoiding explicit dense matrix operations include nullspace projection, deflation, and Sherman-Morrison-Woodbury-based operator reformulation (Wang et al., 2022, Ojeda-Ruiz et al., 22 Oct 2025).
4. Randomization and Distributional Approaches Where feasible sets under strict constraints would be small or empty, randomization over feasible allocations or matchings is used (e.g., distributional maxmin fairness in ranking, mixture models in randomized subset selection) to ensure ex-ante fairness, frequently yielding Pareto improvements or mitigating individual-level harm (Garcia-Soriano et al., 2021, Tang et al., 2023, Panda et al., 2022).
5. Efficient Discrete/Combinatorial Algorithms For resource allocation, knapsack, matching, or clustering, specialized approximation schemes (FPTAS, PTAS, LP rounding, greedy-maximal) address the inherent hardness induced by group fairness constraints, often with provable bicriteria or additive violation guarantees (Patel et al., 2020, Dickerson et al., 2023, Sankar et al., 2021).
6. Post-Training Adjustments/Corrections In statistical learning, group mean constraints can be enforced post hoc via closed-form adjustments to tree/ensemble outputs (e.g., leafwise mean correction in regression forests) or via “noiseless observation” augmentation in kernel regression (Fitzsimons et al., 2018, Fitzsimons et al., 2018).
7. Specific Modern Techniques
   • GraphGini employs a Gini coefficient–based proxy and Nash social welfare–style group parity loss in GNNs (Sirohi et al., 2024).
   • Multi-task RL fairness algorithms check (and explicitly enforce) per-task group return gaps ≤ε across all groups and tasks, using exploration bonuses and feasible policy sets (Song et al., 10 Mar 2025).

4. Theoretical Guarantees and Computational Complexity

Enforcing group fairness constraints almost always increases computational complexity compared to unconstrained settings; indeed, many problem variants are NP-hard or inapproximable without relaxation (Patel et al., 2020, Sankar et al., 2021, Dickerson et al., 2023). Nonetheless, a variety of theoretical guarantees have been established:

• Approximation and Bicriteria Results: Nearly all resource allocation settings provide PTAS/FPTAS or bicriteria approximations that guarantee objective values close to optimal, with fairness constraints met within geometric or additive slack (Patel et al., 2020, Tang et al., 2023, Panda et al., 2022, Dickerson et al., 2023).
• Fairness–Utility Trade-Off Bounds: Quantitative studies consistently report that group fairness can be enforced at moderate loss in utility, with relative reductions in group disparities typically far exceeding the reductions in overall performance (commonly <1–5% utility loss; (Sirohi et al., 2024, Song et al., 10 Mar 2025, Freund et al., 2023, Dickerson et al., 2023)).
• Non-Violation Guarantees: Especially in episodic RL and online matching, sublinear regret and high-probability guarantees on group fairness violations are achieved (Song et al., 10 Mar 2025, Deng et al., 2022, Freund et al., 2023).
• Scalability: Techniques such as nullspace projection and SMW-based operator formulation permit group fairness–constrained spectral clustering to scale comparably to unconstrained methods (Wang et al., 2022, Ojeda-Ruiz et al., 22 Oct 2025).
• Pareto and Nash Welfare Properties: Algorithms incorporating Nash social welfare constraints, as in GraphGini, yield Pareto-optimal trade-offs among groups (Sirohi et al., 2024).

5. Representative Applications Across Domains

Group fairness constraints have been instantiated and analyzed in:

• Graph Learning: GraphGini imposes equal group-wise Gini coefficients on node embeddings, achieving state-of-the-art trade-offs on group and individual fairness in node classification (Sirohi et al., 2024).
• Reinforcement Learning: Both episodic and multi-task RL settings apply constraints to ensure demographic parity of returns or stepwise action distributions (Song et al., 10 Mar 2025, Deng et al., 2022).
• Clustering: Spectral clustering (s-FairSC, Fair-SMW) and $k$-center/median clustering algorithms enforce proportional representation per cluster with explicit linear constraints and scalable eigensolvers (Wang et al., 2022, Ojeda-Ruiz et al., 22 Oct 2025, Dickerson et al., 2023).
• Optimal Transport: FairSinkhorn and penalized OT enforce target pairwise group matchings, enabling exploration of fairness–cost frontiers in matching plans (Bleistein et al., 12 Jan 2026).
• Allocation Problems: Knapsack, subset selection, participatory budgeting, and matching settings use group quotas and per-group value bounds, with randomized and deterministic approximation guarantees (Patel et al., 2020, Sankar et al., 2021, Tang et al., 2023).
• Ranking and Exposure: Distributional maxmin-fair ranking solves for randomized orderings subject to group prefix-exposure constraints, optimizing ex-ante individual exposure (Garcia-Soriano et al., 2021).
• Regression and Tree Models: Post-hoc group mean adjustment ensures expectation constraints across all groups or intersections, with negligible computational overhead (Fitzsimons et al., 2018, Fitzsimons et al., 2018).
• Dynamic Assignment: Bid-price controls and online algorithms for refugee assignment enforce per-group minimum utility constraints, with vanishing regret and interpretable reserve systems (Freund et al., 2023).

6. Practical Implications and Empirical Insights

Extensive empirical analyses establish several robust phenomena:

• Enforcement yields large disparity reductions at small efficiency cost: In node classification, RL, matching, and clustering, group fairness can often be satisfied with <5% reduction in aggregate utility while driving gaps between group outcomes near zero (Sirohi et al., 2024, Song et al., 10 Mar 2025, Freund et al., 2023, Wang et al., 2022, Ojeda-Ruiz et al., 22 Oct 2025).
• Strong theoretical–practical alignment: Runtime and scalability advances (e.g., SMW-based fair spectral clustering) achieve order-of-magnitude speedups while matching unconstrained performance and fairness (Ojeda-Ruiz et al., 22 Oct 2025).
• Flexible constraint specification: Modern systems (e.g., OmniFair) support declarative, model-agnostic specification and joint enforcement of multiple group fairness constraints, covering statistical parity, equalized odds, predictive parity, and user-customized metrics (Zhang et al., 2021).
• Intersectional and multi-group fairness: Post-training correction and expectation constraint approaches enable exact, low-overhead enforcement of fairness across arbitrary intersections of protected attributes (Fitzsimons et al., 2018).
• Analysis of fairness–accuracy trade-offs: Empirical studies consistently show that aggressive group fairness enforcement induces monotonic, marginal accuracy cost curves, supporting Pareto analysis, and facilitating practical policy trade-offs (Sirohi et al., 2024, Freund et al., 2023, Zhang et al., 2021).

7. Limitations and Future Research Directions

Despite significant algorithmic and empirical advances, several open problems and limitations remain:

• Scalability to Many Groups and Constraints: Complexity grows with the number of constraints (notably for intersectionality and multi-task RL), motivating further research into efficient algorithms and approximate feasibility for large-scale, high-dimensional group structures (Song et al., 10 Mar 2025, Fitzsimons et al., 2018).
• Continuous and Soft Fairness Metrics: Most frameworks target hard, discrete constraints (parity, boundaries), with limited support for soft metrics such as equalized odds or continuous relaxations; extensions are suggested but not fully developed (Song et al., 10 Mar 2025).
• Robustness to Distributional Shift and Data Scarcity: Statistical guarantees depend on concentration assumptions that can fail for small or under-sampled groups; adaptive methods are needed to ensure robustness (Freund et al., 2023).
• Tension with Individual Fairness: In general, group fairness constraints are not sufficient to guarantee individual-level equity, and in some cases may introduce new disparities. Recent work explores randomized distributions over feasible allocations to mitigate individual-level harm within group fairness envelopes (Garcia-Soriano et al., 2021, Panda et al., 2022).
• Theoretical Tightness and Price-of-Fairness Bounds: Quantitative characterization of the minimum efficiency sacrifice necessary to achieve arbitrary group fairness constraints remains an open area, particularly in complex combinatorial settings (Tsang et al., 2019, Dickerson et al., 2023).
• Flexible/User-Customizable Specifications: Emerging declarative and programmable fairness systems (OmniFair) promote broad adoption but raise new research questions regarding interpretability, constraint selection, and legal alignment (Zhang et al., 2021).

Group fairness constraints constitute a foundational and technically diverse pillar of modern algorithmic fairness, with an active research landscape dedicated to their mathematical formalization, computational enforcement, practical evaluation, and integration with complementary fairness paradigms.