Evolved Sample Weights for Bias Mitigation: Effectiveness Depends on the Fairness Objective

1. Problem Statement

Machine learning models trained on real-world data often inherit and perpetuate biases (e.g., racial, gender) present in the training data, leading to unfair predictions that negatively impact marginalized communities. While reweighting (assigning specific weights to data points during training) is a common pre-processing technique to mitigate this bias, it often creates a trade-off: improving fairness can degrade predictive accuracy.

Existing approaches to calculating sample weights generally fall into two categories:

1. Deterministic Reweighting: Deriving weights directly from dataset statistics (e.g., Kamiran & Calders).
2. Equal Weights: Assigning a uniform weight of 1.0 to all data points (the baseline).

There is a lack of systematic understanding regarding whether evolutionary approaches (specifically Genetic Algorithms) for learning sample weights can outperform deterministic methods in balancing the trade-off between predictive performance and fairness across different metric combinations.

2. Methodology

A. Experimental Design

The authors conducted a comprehensive evaluation across 11 publicly available datasets (including two medical datasets regarding perinatal mood and anxiety disorders). They compared three sample weighting strategies:

1. Equal Weights (EQ): All data points weighted equally.
2. Deterministic Weights (DW): Weights calculated using a formula based on sensitive attributes and target variable probabilities (Kamiran & Calders).
3. Evolved Weights (EW): Weights optimized using a Genetic Algorithm (GA).

B. The Genetic Algorithm (GA) Approach

The GA is designed to evolve a set of sample weights to simultaneously optimize two objectives:

• Maximize Predictive Performance: Measured by Accuracy (ACC) or Area Under the ROC Curve (AUROC).
• Minimize Fairness Disparity: Measured by Demographic Parity (DP) or Subgroup False Negative Fairness (SFN).

Key GA Mechanics:

• Representation: The "individual" in the population is a vector of weights. The dimensionality corresponds to the unique combinations of sensitive attributes and target variable values (not the total number of samples), making the search space efficient.
• Evolution: Uses NSGA-II (Non-dominated Sorting Genetic Algorithm II).
  • Selection: Non-dominated binary tournament selection based on Pareto rank and crowding distance.
  • Crossover & Mutation: Standard operators with probabilities of 0.8 and 0.1, respectively.
  • Evaluation: Models are trained using 10-fold cross-validation to evaluate both predictive and fairness metrics on the validation set.
• Output: A Pareto front representing the best trade-offs between accuracy and fairness found during the search.

C. Evaluation Metrics

The study evaluated the quality of the trade-offs using Hypervolume, a metric that quantifies the volume of the objective space dominated by a Pareto front relative to a reference point. A larger hypervolume indicates a better set of solutions (more diverse and closer to the true optimum).

The study tested four combinations of objectives:

1. (ACC, DP)
2. (ACC, SFN)
3. (ROC, DP)
4. (ROC, SFN)

D. Statistical Analysis

• Non-hierarchical: Friedman test followed by Wilcoxon signed-rank tests to compare hypervolumes across methods for specific datasets.
• Hierarchical: Bayesian mixed-effects modeling (Zero-One Inflated Beta regression) to analyze global trends and interactions between weighting methods, predictive objectives, and fairness objectives.

3. Key Contributions

1. Systematic Comparison: The first systematic comparison of evolved sample weights (GA) against deterministic and equal weighting methods across multiple predictive/fairness metric combinations.
2. Metric-Dependent Insights: The discovery that the effectiveness of evolved weights is highly dependent on the choice of the fairness objective.
3. Pareto Optimization: Demonstrating that GA-based reweighting can generate Pareto fronts with significantly larger hypervolumes (better trade-offs) than deterministic methods, particularly when optimizing for Demographic Parity.
4. Medical Validation: Successful application and validation of the method on real-world medical datasets (PMAD screening), proving its utility in high-stakes healthcare scenarios.

4. Results

A. Overall Performance

• Evolved Weights (EW) consistently outperformed Equal Weights (EQ) and Deterministic Weights (DW) in terms of hypervolume across the majority of datasets and metric combinations.
• In the (Accuracy, Demographic Parity) condition, EW yielded significantly better results on 10 out of 11 datasets.
• Statistical modeling confirmed a significant main effect of the weighting method on hypervolume.

B. The Role of the Fairness Objective

The most critical finding is that the magnitude of improvement depends heavily on the fairness metric used:

• Demographic Parity (DP): EW showed the most significant advantage over baselines when optimizing for DP.
• Subgroup False Negative Fairness (SFN): The advantage of EW was less pronounced or inconsistent compared to DP.
• Interaction Effect: The statistical model revealed a significant interaction between the weighting method and the fairness objective. The benefit of using a GA is not uniform; it is maximized when the fairness objective is DP.

C. The "Degenerate Strategy" Warning

The authors observed that DP can sometimes be improved by uniformly worsening predictions for all groups (reducing the acceptance rate for everyone) without changing the relative disparity between groups. This can artificially inflate fairness scores while degrading actual utility. The GA successfully navigated this, but the authors warn that metric selection dictates what kind of fairness improvement is encouraged.

5. Significance and Conclusion

Significance:

• Algorithmic Fairness: The paper establishes that evolutionary algorithms are a superior method for finding optimal sample weights compared to static, deterministic formulas, provided the computational cost is acceptable.
• Practical Guidance: It provides a crucial warning to practitioners: The choice of fairness metric is as important as the mitigation technique. Optimizing for Demographic Parity yields different (and often easier) gains than optimizing for False Negative fairness.
• Healthcare Application: The successful application to medical data suggests that evolved reweighting can help reduce health disparities without sacrificing diagnostic accuracy, a vital requirement for clinical AI.

Limitations & Future Work:

• Computational Cost: The GA approach requires significantly more model training calls than deterministic methods.
• Metric Sensitivity: The results highlight that some fairness metrics (like DP) may be susceptible to "gaming" or degenerate solutions. Future work should investigate whether metric improvements translate to genuine real-world benefits for marginalized groups.
• Generalizability: Future research should test these methods on different model architectures and explore alternative GA configurations (e.g., different selection strategies).

In summary, the paper argues that evolved sample weights are a powerful tool for bias mitigation, but their efficacy is not absolute; it is inextricably linked to the specific fairness definition chosen for optimization.