🤖 machine learning

End-to-End Fairness Optimization with Fair Decision-Focused Learning

This paper introduces End-to-End Fairness Optimization (E2EFO) and its associated Fair Decision-Focused Learning (FDFL) framework, which jointly optimizes prediction accuracy, prediction fairness, and decision fairness in resource allocation by training predictors to minimize decision regret through differentiable optimization layers.

Original authors: Yu Wang (Xinying), Violet (Xinying), Chen

Published 2026-08-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yu Wang (Xinying), Violet (Xinying), Chen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the captain of a massive spaceship, and your job is to distribute a limited supply of fuel to different colonies. To do this fairly, you need two things: a crystal ball to predict how much fuel each colony actually needs, and a rulebook for how to hand out the fuel once you have those predictions. This is the world of "prediction-informed decision-making," a field where computers try to guess the future so humans can make better choices. But here's the tricky part: what if your crystal ball is slightly biased? Maybe it's better at guessing the needs of the rich colonies than the poor ones. If you use a biased crystal ball to follow a fair rulebook, you might accidentally end up being unfair to everyone. This paper tackles that exact headache. It asks a simple but profound question: How do we fix the crystal ball and the rulebook at the same time, so that the final outcome is truly fair for everyone, no matter their background?

The authors of this paper, Yu Wang and Violet Chen, propose a new way to train computer models called "End-to-End Fairness Optimization" (E2EFO). Think of it as a "fairness gym" for artificial intelligence. Usually, when we train an AI, we tell it to be as accurate as possible at guessing the future (the prediction stage) and then we use those guesses to make decisions. But the authors realized that being accurate isn't enough. If the AI makes small, unfair mistakes for certain groups of people, those mistakes can snowball into huge unfairness when the AI tries to distribute resources like healthcare or money.

To solve this, they introduced a training method called "Fair Decision-Focused Learning" (FDFL). Imagine you are teaching a student to bake a cake. A traditional teacher might say, "Just make sure your ingredients are measured perfectly" (prediction accuracy). But FDFL is like a teacher who says, "Measure your ingredients, but also make sure you don't accidentally leave out the sugar for the gluten-free guests, and remember, the final cake needs to taste good for everyone." This method trains the AI to care about three things at once:

Accuracy: Is the guess close to the truth?
Prediction Fairness: Does the AI make the same kind of mistakes for everyone, or does it mess up more for one specific group?
Decision Fairness: When the AI uses its guesses to hand out resources, does the final result feel fair to all the groups involved?

The paper finds that you can't just fix one part and ignore the others. In their experiments, which included simulating how to distribute healthcare resources to patients and how to allocate multiple types of resources to different groups, they showed that training the AI to look at all three goals together works best. They discovered that if the AI is not very smart (a "low-capacity" model) or if the rules for fairness are very strict, ignoring the "decision fairness" part leads to bad outcomes. Conversely, if the data is heavily biased against certain groups, ignoring the "prediction fairness" part leads to the AI learning those biases.

The researchers didn't just guess this; they ran thousands of simulations. They found that their new method, FDFL, consistently reduced the "regret"—which is a fancy word for "how much worse the outcome was because the AI wasn't perfect"—compared to older methods. They also proved mathematically that their method is stable and won't just work on the practice data but should work on new, unseen data too. The key takeaway is that fairness isn't a single switch you flip; it's a delicate balance. By teaching the AI to juggle accuracy, prediction fairness, and decision fairness all at once, we can build systems that don't just look good on paper but actually do the right thing in the real world.

Technical Summary: End-to-End Fairness Optimization with Fair Decision-Focused Learning

1. Problem Formulation

The paper addresses the challenge of End-to-End Fairness Optimization (E2EFO) in data-driven decision-making systems where predictive models inform resource allocation decisions. The authors argue that fairness must be addressed jointly across two distinct stages:

Prediction Stage: Estimating unknown impacts (e.g., health benefits, creditworthiness) from historical data. Fairness here concerns prediction disparity, ensuring that prediction errors do not systematically disadvantage specific groups.
Decision Stage: Allocating limited resources based on predictions to maximize equitable outcomes. Fairness here concerns decision regret, the loss in decision quality (equity) caused by relying on imperfect predictions rather than ground-truth parameters.

The authors illustrate that optimizing for fairness in only one stage can lead to unfair outcomes in the other. For instance, a fair predictor with noisy estimates might lead a decision-maker to over-allocate resources to a disadvantaged group, creating unfairness against an advantaged group. Conversely, a highly accurate but biased predictor might lead to equitable-looking allocations that are actually unjust.

The specific problem is formalized as a resource allocation task where:

Stakeholders are partitioned into groups.
Prediction: A model $f_\theta$ estimates impacts $\hat{r}$ from features $X$ .
Decision: An optimization model allocates resources $d$ to maximize a group-based $\alpha$ -fairness utility function $W_\alpha^g(u)$ , subject to convex constraints (e.g., budget).
Objective: Minimize three metrics simultaneously:
1. Prediction Accuracy: Mean Squared Error (MSE).
2. Prediction Fairness: Mean Absolute Deviation (MAD) of group-level MSEs.
3. Decision Regret: The gap between the fairness score achieved using predictions and the optimal fairness score achievable with true impacts.

2. Methodology: Fair Decision-Focused Learning (FDFL)

The paper proposes Fair Decision-Focused Learning (FDFL), a training paradigm that treats E2EFO as a Multi-Task Learning (MTL) problem. Unlike standard Predict-Then-Optimize (PTO) or standard Decision-Focused Learning (DFL) which optimize single objectives, FDFL jointly optimizes the three conflicting objectives.

Key Technical Components:

Gradient-Based Training: The predictor $f_\theta$ $f_{θ}$ is trained via gradient descent. The gradients for the three objectives are computed and combined:
- $\nabla_\theta L_{pred}$ and $\nabla_\theta F$ (accuracy and fairness) are standard backpropagation.
- $\nabla_\theta L_{regret}$ requires differentiating through the decision optimization problem.
Decision Jacobian Computation: A core computational challenge is computing $\frac{\partial d^*(\hat{r})}{\partial \hat{r}}$ $\frac{\partial d ^{*} ( r ^ )}{\partial r ^}$ . The authors provide two solutions:
1. Closed-Form: For single-budget $\alpha$ -fair allocation (knapsack problem), they derive exact closed-form formulas for the optimal solution and its Jacobian using Karush-Kuhn-Tucker (KKT) conditions.
2. Differentiable Optimization: For general convex feasible sets, they employ differentiable convex optimization layers (specifically cvxpylayers) to implicitly differentiate the optimality conditions.
Gradient Combination Strategies: To handle the multi-objective nature, FDFL employs MTL techniques to combine gradients into a single update direction:
- Static Scalarization (FDFL-Scal): A fixed weighted sum of the three objectives.
- Dynamic Conflict-Avoidant Methods: Including PCGrad (projecting conflicting gradients onto normal planes) and Nash-MTL (Nash bargaining weights).
Theoretical Guarantees: The paper establishes a finite-sample generalization bound for the scalarized FDFL objective. Under boundedness and Lipschitz regularity assumptions, the composite excess risk of the empirical minimizer is shown to be $O(\sqrt{q/N})$ , where $q$ is the parameter count and $N$ is the sample size.

3. Key Contributions

The authors summarize their contributions as follows:

Framework: Introduction of E2EFO as a unifying framework integrating group-based prediction fairness (MAD of errors) and group-based decision fairness (two-level $\alpha$ -fairness) into a single optimization pipeline.
Algorithm: Development of FDFL algorithms that utilize MTL techniques (static and dynamic) to train predictors balancing accuracy, prediction fairness, and decision regret. This includes the derivation of exact closed-form decision Jacobians for specific allocation tasks and the application of differentiable optimization layers for general cases.
Theory: Establishment of a finite-sample generalization bound for scalarized FDFL training, verifying the necessary conditions for $\alpha$ -fairness allocation tasks.
Empirical Validation: Evaluation on a healthcare-based single-resource allocation and a synthetic multiple-resource allocation, demonstrating the value of joint optimization.

4. Experimental Results

The authors evaluated FDFL against baselines including PTO, standard DFL, and fairness-embedded PTO variants across varying predictor capacities (linear, MLP-16, MLP-64), group imbalances, and fairness parameters ( $\alpha$ ).

Key Findings:

Multi-Objective Necessity: Training on a single objective (e.g., regret only) often degrades other metrics. For example, DFL (regret-only) achieved the lowest regret but the highest prediction disparity. Joint training (FDFL) successfully reduced disparity without sacrificing decision quality.
Complementary Roles:
- Prediction Fairness is most effective at reducing disparity when group imbalance is high.
- Decision Regret optimization provides the most significant gains when the predictor capacity is limited (leading to larger errors) or when the decision fairness parameter $\alpha$ is large (making the allocation highly sensitive to prediction errors).
Predictor Capacity: In settings with high-capacity predictors (MLP-64), prediction accuracy and decision quality were often aligned, reducing the gap between PTO and DFL. However, with lower capacity, the decision-regret objective became crucial to mitigate error propagation.
Robustness: The proposed FDFL methods consistently outperformed baselines across different imbalance levels and fairness parameters, particularly when all three objectives were included in the training loss.

5. Significance and Claims

The paper claims to fill a persistent gap in the literature by providing an integrated framework that addresses fairness in both the prediction and decision stages simultaneously, rather than treating them in isolation.

Theoretical Significance: It extends the theoretical understanding of decision-focused learning to include fairness constraints and provides the first generalization bounds for such multi-objective E2EFO settings.
Practical Significance: The results demonstrate that "fair predictions" do not guarantee "fair decisions" and vice versa. The authors argue that practitioners must adopt a holistic training approach (FDFL) to ensure fairness throughout the entire pipeline, especially in critical domains like healthcare and social services where systemic biases and resource scarcity coexist.
Modesty: The authors acknowledge limitations, noting that their generalization bound is specific to fixed scalarization weights and may become vacuous in highly over-parameterized regimes ( $q > N$ ). They also leave the extension of E2EFO to stochastic decision views (where predictions are distributions rather than point estimates) for future work.