Fairness Aware Automl Eval

fairness-aware-automl-eval

Exploring the impact of fairness-aware criteria in AutoML — Simões and Correia (2026) (arXiv:2604.10224, 2026)

What this evaluates

This evaluation probes an AutoML framework's ability to jointly optimize predictive performance and fairness constraints during pipeline search. It measures how well a multi-criteria genetic algorithm balances accuracy metrics against demographic parity, equalised odds, and ABROCA while simultaneously selecting data and features.

Datasets

• adult — total 48842; splits: test (-1)
• credit-card — total 30000; splits: test (-1)
• portuguese-bank-marketing — total 45211; splits: test (-1)

Metrics

• DP (primary) — range: [0, 1]
  • Demographic Parity measures the statistical parity of positive prediction rates across sensitive groups. All metrics are transformed and scaled to [0, 1] for the optimization objective, where 0 represents the optimal value (minimisation of unfairness).
• EO — range: [0, 1]
  • Equalised Odds measures the equality of true positive and false positive rates across sensitive groups. Scaled to [0, 1] where 0 is optimal.
• ABROCA — range: [0, 1]
  • Area Between ROC Curves evaluates fairness via ROC curve area differences across sensitive slices. Scaled to [0, 1] where 0 is optimal.
• MCC — range: [0, 1]
  • Matthews Correlation Coefficient measures predictive performance. Scaled to [0, 1] where 0 is optimal after transformation.
• TPR — range: [0, 1]
  • True Positive Rate measures predictive performance. Scaled to [0, 1] where 0 is optimal after transformation.

Input / output format

Input: Binary classification datasets containing sensitive attributes. The AutoML framework receives stratified training folds and optimizes a pipeline (data selection + model tuning) under a 30-minute time budget.

Output: The best-performing pipeline configuration and its corresponding performance/fairness scores evaluated on the held-out test set.

Scoring recipe

# Normalize to [0,1] where 0 is optimal (minimization)
def scale(val, higher_better=True):
    return 1.0 - val if higher_better else val

# Compute on test set
dp = compute_demographic_parity(y_true, y_pred, sensitive)
eo = compute_equalised_odds(y_true, y_pred, sensitive)
abroca = compute_abroca(y_true, y_pred, sensitive)
mcc = compute_mcc(y_true, y_pred)
tpr = compute_tpr(y_true, y_pred)

# Scaled values for optimization
dp_s, eo_s, abroca_s = scale(dp), scale(eo), scale(abroca)
mcc_s, tpr_s = scale(mcc, higher_better=True), scale(tpr, higher_better=True)

# Fitness function (alpha=0.8)
alpha = 0.8
fitness = alpha * (dp_s + eo_s + abroca_s) / 3 + (1 - alpha) * (mcc_s + tpr_s) / 2
return fitness

Common pitfalls

• All reported metrics are transformed and scaled to [0, 1] for the optimization objective, where 0 represents the optimal value (minimisation). Readers must account for this inversion when comparing to standard metric conventions.
• Results are averaged across 30 independent runs with 5-fold cross-validation, yielding 150 solutions per dataset. Statistical significance is assessed using t-tests or Wilcoxon signed-rank tests with Bonferroni correction.
• The framework optimizes both data selection (instance/feature) and model tuning simultaneously, meaning reported performance includes the effects of data reduction, not just model accuracy.

Evidence (verbatim from paper)

All the metrics are normalised and scaled to [0, 1], where 0 represents the optimal value for our minimisation objective of reducing both predictive error and unfairness. Appropriate transformations were applied to convert all metrics to this scale, specially Equation [3] which originally uses a different range of values. DP and EO fairness metrics were calculated using the Fairlearn framework, while ABROCA was implemented based on Mangal et al. work.

Citation

@misc{simoes2026fairnessawareautoml,
  title={Exploring the impact of fairness-aware criteria in AutoML},
  author={Simões and Correia (2026)},
  year={2026},
  note={arXiv:2604.10224}
}

• arXiv: 2604.10224