Fairness Repair Eval

fairness-repair-eval

Fix Fairness, Don't Ruin Accuracy: Performance Aware Fairness Repair using AutoML — Giang Nguyen et al. (2023) (arXiv:2306.09297, 2023)

What this evaluates

Evaluates the ability of an AutoML-based fairness repair framework to mitigate bias in machine learning models while preserving predictive accuracy. It measures the trade-off between accuracy retention and bias reduction across multiple binary classification datasets and model architectures.

Datasets

• Adult Census (race) — total 32561; splits: test (-1)
• Bank Marketing (age) — total 41188; splits: test (-1)
• German Credit (sex) — total 1000; splits: test (-1)
• Titanic (sex) — total 891; splits: test (-1)

Metrics

• Accuracy difference (primary) — range: other
  • new accuracy - old accuracy. Positive values indicate improved predictive performance after repair.
• DI difference — range: other
  • old DI - new DI. Measures reduction in Disparate Impact bias.
• SPD difference — range: other
  • old SPD - new SPD. Measures reduction in Statistical Parity Difference bias.
• EOD difference — range: other
  • old EOD - new EOD. Measures reduction in Equalized Odds Difference bias.
• AOD difference — range: other
  • old AOD - new AOD. Measures reduction in Average Odds Difference bias.

Input / output format

Input: Binary classification datasets containing feature columns, target labels, and a designated sensitive attribute column (race, age, or sex). Models are trained on these datasets to produce baseline predictions before repair.

Output: Repaired model predictions or optimized hyperparameter configurations generated by the Fair-AutoML framework, evaluated against the original baseline predictions.

Scoring recipe

def compute_evaluation(original_preds, repaired_preds, labels, sensitive_attr):
    acc_diff = accuracy_score(labels, repaired_preds) - accuracy_score(labels, original_preds)
    bias_diff = compute_fairness_metric(original_preds, sensitive_attr) - compute_fairness_metric(repaired_preds, sensitive_attr)
    return acc_diff, bias_diff
# Positive acc_diff and bias_diff indicate improvement in accuracy and fairness respectively.

Common pitfalls

• Misinterpreting the sign of the difference metric: in this protocol, positive values indicate improvement for both accuracy and fairness, which is the opposite of standard loss-based metrics.
• Failing to use the correct sensitive attribute per dataset (e.g., using 'sex' for Adult Census instead of 'race'), leading to incorrect fairness calculations.
• Not distinguishing between the four fairness variants (DI, SPD, EOD, AOD), as they capture different bias dimensions and may yield conflicting results.

Evidence (verbatim from paper)

For accuracy, accuracy difference = new accuracy - old accuracy. For bias (DI, SPD, EOD, AOD), bias difference = old bias - new bias. Thus, a positive value indicates an improvement in bias/accuracy in the repaired model compared to the original and vice versa.

Citation

@misc{nguyen2023fixfairness,
  title={Fix Fairness, Don't Ruin Accuracy: Performance Aware Fairness Repair using AutoML},
  author={Giang Nguyen et al. (2023)},
  year={2023},
  note={arXiv:2306.09297}
}

• arXiv: 2306.09297