Abstract
Machine learning systems are often trained using data collected from historical decisions. If past decisions were biased, then automated systems that learn from historical data will also be biased. We propose a black-box approach to identify and remove biased training data. Machine learning models trained on such debiased data (a subset of the original training data) have low individual discrimination, often 0%. These models also have greater accuracy and lower statistical disparity than models trained on the full historical data. We evaluated our methodology in experiments using 6 real-world datasets. Our approach outperformed seven previous approaches in terms of individual discrimination and accuracy.
Introduction
Machine learning models are only as good as the data they consume. When we train models on historical decisions—such as loan approvals or parole hearings—we often bake in human prejudices. This manifests as label bias (biased human graders) or selection bias (unrepresentative sampling).
The standard industry response has been to treat fairness as a constraint, forcing the model to be fair during training or fixing its outputs afterward. However, this usually creates a "fairness-accuracy tradeoff," where the model becomes less useful as it becomes more equitable. This paper challenges that tradeoff by asking: What if we just stop feeding the model the "garbage" data that causes the bias in the first place?
Methodology
The authors propose a three-step pre-processing loop that treats the machine learning model as a "black box," meaning it works regardless of the specific architecture (neural nets, decision trees, etc.).
- Identify Unfair Decisions: The system generates "similar pairs"—synthetic individuals who are identical in every way except for a sensitive attribute (like race or sex). If the model gives them different outcomes, it is flagged as a "discriminatory pair."
- Trace the Influence: Using influence functions, the researchers trace these unfair predictions back to the specific training data points that most heavily influenced them. The hypothesis is that the training points most responsible for discriminatory behavior are the most biased.
- Prune and Retrain: The most "influential" biased points are removed from the dataset. The model is then retrained on this "debiased" subset.
Experimental Results
The methodology was evaluated using 8 experiments across 6 well-known real-world datasets (e.g., Adult Income, German Credit, COMPAS Recidivism). The authors compared their approach against seven baseline techniques, including "Massaging" (MA), "Adversarial Debiasing" (AD), and "Learning Fair Representations" (LFR).
The "Win-Win" Discovery
The most striking finding is that removing biased data often increases test accuracy. For example, on the Adult Income dataset (D1), the debiased model reached 92-93% accuracy, while the model trained on the full (biased) dataset hovered around 80-82%. Biased data points act like noise or outliers; by removing them, the model learns a more "principled" and generalizable logic rather than memorizing past prejudices.
Fairness Metrics
- Individual Discrimination: The technique achieved nearly 0% individual discrimination across all datasets.
- Group Fairness: While the paper prioritizes individual discrimination (similar people getting similar results), solving for the individual often naturally improved group-level statistics like statistical parity difference.
Limitations and Future Work
Despite the strong results, several hurdles remain. Generating synthetic "similar individuals" is difficult for complex datasets (e.g., handling categorical attributes like "college attended" realistically). Furthermore, the algorithm requires access to sensitive attributes during the cleaning phase, which may be legally restricted in some environments. Finally, calculating influence functions for massive datasets can be computationally expensive, representing an area for future optimization.
Summary updated