From a technical perspective, fairness definitions can be mutually incompatible under realistic conditions [2– 6]. Calibration, Equalized Odds (EO), and Demographic Parity (DP) cannot, in general, be simultaneously satisfied when base rates differ across groups. Additionally, interpretability constrains the hypothesis class: Restricting models to linear or rule-based forms limits the function space that can be explored. From an organizational perspective, ML teams negotiate among stakeholders with different priorities, sometimes legally mandated. The EU Artificial Intelligence (AI) Act [7] imposes transparency and fairness obligations on high-risk systems, while US fair lending laws require adverse-action explanations. Without an explicit multi-objective formulation, trade-offs are often made implicitly by the engineer who selects the loss function or threshold, without consulting affected communities or documenting the reasoning.

This paper is organized as follows. Section 2 formalizes the MOO framework. Section 3 reviews fairness definitions and metrics, including key impossibility results. Section 4 addresses interpretability concepts, methods, and pitfalls. Section 5 surveys MOO methods applicable to ML, including scalarization, constrained learning, evolutionary algorithms, and multi-gradient approaches. Section 6 proposes an evaluation and reporting framework with visualization tools, metric selection guidance, and a documentation checklist, and Section 7 concludes and identifies directions for future work.

Let A denote a binary sensitive attribute (e.g., race, gender), Y the true label, Ŷ the predicted label, and S a score output by the classifier. The following group fairness definitions are widely used:

I. DP: P(Ŷ = 1 | A = 0) = P(Ŷ = 1 | A = 1). The positive classification rate should be equal across groups, regardless of the true label distribution.

II. EO: P(Ŷ = 1 | Y = y, A = a) is equal across groups for y ∈ {0, 1}. Both true positive rates and false positive rates should be equal across groups [1].

III. Calibration: P(Y = 1 | S = s, A = a) = s for all groups. Among individuals assigned scores, the proportion who are truly positive should equal s, regardless of group membership [6].

Impossibility results: A fundamental finding in the fairness literature is that, except in degenerate cases (equal base rates or perfect prediction), DP, EO, and calibration cannot all be satisfied simultaneously [2], [5]. This impossibility result is not a limitation of any particular algorithm but a mathematical consequence of the structure of the problem. It implies that any real-world system must choose among fairness criteria a normative decision that cannot be resolved by technical means alone.

Individual fairness [9] requires that similar individuals receive similar predictions: if d(x, x') is a task-specific similarity metric over individuals, then |f(x) − f(x')| should be small whenever d(x, x') is small. The practical challenge lies in defining the metric d, which encodes substantive judgments about which differences between individuals are morally relevant.

Counterfactual fairness requires that a prediction remain unchanged under a hypothetical intervention that changes the sensitive attribute. Formally, P(ŶA←a | X = x, A = a) = P(ŶA←a' | X = x, A = a) for all a, a'. This argument requires a causal model of the data-generating process, which may be unavailable or contested.

Enforcing group fairness constraints typically reduces overall predictive accuracy, because the unconstrained optimum generally violates fairness criteria when base rates differ across groups. The Pareto front between accuracy and fairness violation reveals a characteristic structure: small improvements in fairness near the unconstrained optimum are "cheap" in terms of accuracy loss, while achieving near-parity becomes increasingly expensive as the fairness constraint tightens. This diminishing-returns structure has been empirically observed across datasets and fairness criteria [10], [8] and has important implications for practice: Organizations can often achieve meaningful fairness improvements with modest accuracy costs, but eliminating the last residual disparity may require substantial sacrifices.

Lipton [11] distinguishes between transparency, the degree to which a model's internal mechanisms are understandable to a human, and post-hoc explainability, the degree to which a separate method can provide insight into a model's behavior after training. Transparency is a property of the model itself; explainability is a property of an explanation method applied to a model.

Inherently interpretable models include linear models, logistic regression, decision trees, rule lists, and generalized additive models. Their structure permits direct inspection of how inputs map to outputs. Post-hoc explanation methods include Local Interpretable Model-agnostic Explanations (LIME) [12], Shapley Additive Explanations (SHAP) [13], attention visualization, saliency maps, and counterfactual explanations. These methods approximate or summarize the behavior of complex models without altering the model itself.

The conventional wisdom holds that simpler, more interpretable models sacrifice predictive accuracy relative to complex black-box models. However, Rudin [3] argues compellingly that for structured and tabular data, inherently interpretable models can match or closely approach the performance of black-box models when sufficient effort is invested in feature engineering and model selection. The accuracy gap is often smaller than assumed, particularly in domains where the underlying relationships are approximately linear or additive. In contrast, for unstructured data (images, natural language, audio), deep neural networks with millions or billions of parameters currently dominate, and interpretable alternatives lag significantly. In these domains, post-hoc explanation methods are the primary tool for understanding model behavior, despite their known limitations.

Several recent findings raise concerns about the reliability of post-hoc explanations:

I. Instability and manipulability: Slack et al. [14] demonstrated that adversarial classifiers can be constructed to produce arbitrary LIME and SHAP explanations while maintaining discriminatory behavior. It shows that post-hoc explanations can be "gamed" to hide unfairness.

II. Attention as explanation: Jain and Wallace [15] showed that attention weights in neural networks do not reliably indicate feature importance. Alternative attention distributions can produce identical predictions, undermining the use of attention as an explanation mechanism.

III. Interpretability theater: There is a risk that post-hoc explanations serve primarily to satisfy regulatory checkboxes or stakeholder expectations, rather than providing a genuine understanding of model behavior. Organizations may deploy explanations that appear transparent but do not actually support meaningful oversight.

These findings reinforce Rudin's [3] argument for preferring inherently interpretable models in high-stakes applications where possible, and for exercising caution when relying on post-hoc methods.

Scalarization converts a multi-objective problem into a single-objective problem by combining objectives into a scalar function.

Weighted sum: Minimize ∑k wk fk(θ), where wk ≥ 0 and ∑ wk = 1. By varying the weight vector, different Pareto-optimal solutions can be recovered. However, linear scalarization can only find solutions on the convex hull of the Pareto front [4]. Solutions in non-convex regions are inaccessible.

Chebyshev scalarization: Minimize maxk wk |fk(θ) − zk*|, where z* is a reference point (e.g., the ideal point). This approach can find solutions on non-convex Pareto regions and provides full coverage of the front. Epsilon-constraint method: Minimize f1(θ) subject to fk(θ) ≤ εk for k = 2, …, K. By systematically varying the ε bounds, the Pareto front can be traced. This method also achieves full Pareto coverage.

A natural approach for fairness-aware ML is to treat fairness as a constraint rather than an objective: minimize L(θ) subject to fairness_metric(θ) ≤ ε. This formulation directly controls the maximum allowable fairness violation while optimizing predictive performance within that bound.

Lagrangian relaxation introduces dual variables λ to convert the constrained problem into an unconstrained saddle-point problem: minθ maxλ ≥ 0 L(θ) + λ. (fairness_metric(θ) − ε). Practical implementations include projected gradient descent and the exponentiated gradient method of Agarwal et al. [10], which reduces fair classification to a sequence of cost-sensitive classification problems. The exponentiated gradient approach is particularly attractive because it can wrap any base classifier and provides convergence guarantees.

Evolutionary multi-objective algorithms maintain a population of solutions and evolve them toward the Pareto front over successive generations.

NSGA-II uses non-dominated sorting to rank solutions by dominance level and crowding distance, thereby maintaining diversity along the front [16]. It is the most widely used evolutionary MOO algorithm and has been applied to hyperparameter optimization, architecture search, and fair model selection.

MOEA/D decomposes the multi-objective problem into a set of scalar subproblems using weight vectors and solves them simultaneously, exploiting neighborhood structure [17]. It is particularly effective for problems with many objectives (> 3).

Evolutionary methods naturally produce diverse, well-distributed Pareto fronts in a single run. Their primary limitation is computational cost: Each generation requires evaluating the entire population, which can be prohibitive for large ML models that require expensive training procedures.

The Multiple-Gradient Descent Algorithm (MGDA) finds a single descent direction that simultaneously reduces all objectives by solving a minimum-norm problem over the convex hull of the individual objective gradients [18]. When such a common descent direction exists, MGDA provides an efficient, gradient-based path toward the Pareto front without requiring weight specification.

Pareto MTL extends this idea to multi-task learning [19]. It generates a set of well-distributed Pareto-optimal solutions by decomposing the multi-objective problem into constrained subproblems, each targeting a different region of the Pareto front. The result is a diverse set of models spanning the trade-off surface, enabling practitioners to select the solution that best aligns with their priorities.

Effective visualization is essential for communicating trade-offs to diverse stakeholders. We recommend the following visualization tools:

I. Pareto front plots (2D and 3D): Scatter plots of objective values for all candidate solutions, with Pareto-optimal solutions highlighted. These are the primary tools for communicating the trade-off structure.

II. Fairness-utility curves: Plots of accuracy (or other utility metrics) against fairness violation for a sweep of constraint thresholds, illustrating the cost of fairness.

III. Radar/spider charts: Multi-metric comparison of selected models on a common scale, useful for presenting multiple objectives simultaneously.

IV. Parallel coordinate plots: For high-dimensional objective spaces (>3 objectives), they allow visualization of relationships among all objectives across the solution set.

Evaluating the quality of a Pareto front approximation requires specialized indicators:

I. Dominance-based indicators: The hypervolume indicator measures the volume of objective space dominated by the Pareto front relative to a reference point. It is the only unary indicator that is strictly monotone with respect to Pareto dominance. The epsilon-indicator measures the minimum factor by which every point in a reference set must be multiplied to be dominated by the approximation set.

II. Distribution indicators: Spacing measures the uniformity of the distribution of solutions along the front.

Spread measures the extent of the front, ensuring solutions cover the full range of trade-offs.

III. Per-group metrics: Beyond aggregate performance, report all relevant metrics disaggregated by protected group: true positive rate, false positive rate, positive predictive value, and calibration error for each group.

We propose the following checklist for reporting multi-objective ML results, designed to promote transparency, reproducibility, and accountability:

Objectives Objective definitions Formal definitions of each objective function, including mathematical formulation and units Objectives Trade-off structure Characterization of conflicts between objectives, including known impossibility results Data Protected attributes List of sensitive attributes and their distributions in training and evaluation data Data Data splits Training/validation/test split strategy, including stratification and temporal considerations Method MOO algorithm Algorithm used and justification for its selection over alternatives

Method Hyperparameters All MOO-specific hyperparameters (weights, constraint thresholds, population sizes, etc.) Results Pareto front Visualization and tabulation of Pareto-optimal solutions with objective values Results Selected solution Final model choice from the Pareto set and rationale for selection

Results Fairness metrics Per-group performance metrics for the selected model and key alternatives Limitations Known gaps Acknowledged limitations, failure modes, and conditions under which results may not generalize

Once the Pareto front has been identified, a decision-maker must select a specific solution for deployment. Several structured approaches exist:

I. Knee-point selection: Identify the "knee" of the Pareto front, the point of maximum marginal rate of substitution, where small improvements in one objective require increasingly large sacrifices in another. This argument represents a natural compromise.

II. Stakeholder preference elicitation: Systematically gather preferences from relevant stakeholders (e.g., domain experts, affected communities, compliance officers) and use these to define weights or aspiration levels. III. Lexicographic ordering: Rank objectives by priority and optimize sequentially: first optimize the most important objective, then optimize the second within the set of solutions that are optimal for the first, and so on.

IV. Satisficing thresholds: Define minimum acceptable levels for each objective and select from solutions that meet all thresholds. If multiple solutions satisfy all constraints, use a secondary criterion (e.g., robustness, simplicity) to select among them.