TL;DR
This paper presents the largest empirical study to date on machine learning bias mitigation, evaluating 17 prominent methods across 8 software decision tasks. Using a comprehensive suite of 11 performance metrics and 4 fairness metrics, the authors reveal that there is no "silver bullet" solution: bias mitigation frequently degrades model performance, and in 25% of cases, it actually leads to a lose-lose outcome where both fairness and performance suffer.
What Problem Does This Paper Address?
As machine learning (ML) increasingly drives critical decisions like hiring, loan approvals, and criminal justice sentencing, software bias has emerged as a major operational concern. Software engineers treat these biases as "fairness bugs" and deploy various bias mitigation methods to fix them.
However, existing evaluations of these mitigation tools are dangerously narrow. Most studies only measure a model's overall accuracy and a single fairness metric. This creates a massive blind spot: in the real world, different applications require different performance trade-offs (such as high precision for a specific minority class in targeted advertising). Without a holistic evaluation, engineers cannot know how a bias mitigation method will impact the specific metrics they care about, or whether the gained fairness is simply an artifact of destroying the model's predictive power.
Methods
The authors conducted a massive, systematic study applying 17 bias mitigation methods to models trained using 3 traditional ML algorithms (Logistic Regression, SVM, Random Forest) and 4 Deep Neural Network (DNN) architectures.
The 17 mitigation methods span three distinct categories:
- Pre-processing: Modifying the training data before training occurs (e.g., Reweighting, Disparate Impact Remover).
- In-processing: Altering the model's learning objective during training to penalize bias (e.g., Adversarial Debiasing, Prejudice Remover).
- Post-processing: Adjusting the prediction outputs of a trained model (e.g., Reject Option Classification).
The study includes 15 methods from the mainstream ML community (via the IBM AIF360 toolkit) and 2 recent methods from the software engineering community (Fairway and Fair-SMOTE). To rigorously evaluate the fairness-performance trade-offs, the authors utilized Fairea, a benchmarking framework that uses model mutation to establish a baseline of naive trade-offs.

Datasets
The evaluation was conducted across 5 widely adopted benchmark datasets, configured into 8 distinct software decision tasks based on the protected attributes being evaluated:
Key Insights and Ideas
- The "Performance Tax" is Pervasive: Applying bias mitigation significantly decreased ML performance in 53% of the studied scenarios.
- The Danger of a Single Metric: Performance drops are highly metric-dependent. A mitigation method might maintain overall accuracy but severely tank the precision or recall for a specific class, proving that single-metric evaluations hide critical failures.
- The Lose-Lose Trap: In 25% of all scenarios, mitigation methods resulted in a "lose-lose" outcome—simultaneously worsening both fairness and performance compared to the original unmitigated model.
- No Universal Winner: No single method consistently achieves the best trade-off. For instance, Reweighting (RW) is excellent at retaining overall ML performance because it only adjusts sample weights, whereas Learning Fair Representations (LFR) maximizes fairness but completely devastates predictive performance.
Key Results
- Fairness vs. Performance: While performance dropped in 53% of scenarios, fairness significantly improved in 46% of scenarios.
- Trade-off Success Rates: According to the Fairea framework, Reweighting (RW) achieved a win-win or good trade-off in the most scenarios (77%). Conversely, Calibrated Equalized Odds with a weighted metric (CEO-W) resulted in a lose-lose trade-off in 68% of cases.
- The Best Overall Performer: Fairway (from the SE community) emerged as the most generally effective tool, achieving the best quantitative trade-off in 30% of the scenarios by intelligently combining pre-processing data filtering with in-processing multi-objective optimization.
Limitations and Future Work
The authors acknowledge two primary limitations in current research:
- Benchmark vs. Production Systems: The study relies on standard open-source benchmarks. Evaluating these tools on real, deployed production systems remains difficult because companies rarely open-source software containing active fairness bugs due to legal and privacy constraints.
- Single Protected Attributes: Most existing tools only mitigate bias for one attribute at a time (e.g., optimizing for sex or race, but not both simultaneously), ignoring intersectional fairness challenges.
Future Directions: The authors highlight the need for hybrid mitigation strategies that cross-cut the pre-, in-, and post-processing phases. They also call for the development of adaptive frameworks that can automatically select the best mitigation strategy based on an application's specific task and target metrics.
Summary updated