fairscope

Subgroup-stratified, calibration-aware fairness auditing for machine-learning models — grounded in peer-reviewed methods.

📖 Documentation: https://rajveer-code.github.io/fairscope/

fairscope packages statistical machinery that mainstream fairness toolkits do not expose as first-class, subgroup-stratified functions, and adds one novel protocol on top:

• DeLong confidence intervals for per-subgroup AUC (fast midrank algorithm).
• Per-subgroup Expected/Maximum Calibration Error with reliability diagrams.
• Significance testing of subgroup performance gaps (paired/unpaired DeLong, stratified bootstrap) with Bonferroni / Benjamini–Hochberg correction.
• A subgroup-stratified interface to standard recalibration — temperature scaling (Guo et al. 2017) and isotonic regression (Zadrozny & Elkan 2002), with pre/post-ECE.
• A novel five-axis Cross-Platform Fairness Evaluation (CPFE) protocol.
• One-call domain audits: healthcare, lending, federated.

Only the CPFE protocol is presented as novel. Every other function ports a documented method and cites its source; the recalibration methods are standard, and the contribution there is the per-subgroup interface and pre/post-ECE reporting.

Status — v0.3.0. All five modules (core, healthcare, nlp/CPFE, federated, lending) are implemented, tested, and released. 100% line coverage on the statistical core; CI green across Python 3.9–3.12. See docs/DESIGN.md for methods and design.

Install

pip install fairscope

Releases are uploaded to PyPI by the maintainer; if a version isn't available there yet, install from source or from the release assets:

git clone https://github.com/Rajveer-code/fairscope
cd fairscope
pip install -e ".[dev]"
pytest

The base install is light (NumPy, SciPy, scikit-learn, pandas, matplotlib). Optional extras: fairscope[nlp] (torch, transformers, captum), fairscope[lending] (econml), fairscope[shap], fairscope[docs].

Quickstart

from fairscope.healthcare import HealthcareFairnessAudit

# y_true : binary outcomes
# y_score: the model's positive-class probabilities
# age_group: a protected attribute, aligned row-for-row
report = HealthcareFairnessAudit.from_scores(
    y_true, y_score, {"age_group": age_group}
).run()

print(report.summary())     # per-subgroup AUC (DeLong CI), ECE, Brier, F1; flags the largest gap
report.to_dataframe()       # tidy per-subgroup table
report.plot_auc_forest()    # forest plot of per-subgroup AUC with DeLong intervals

Every domain is also reachable through one dispatcher, FairnessAudit(model, domain=...), with domain in {"healthcare", "nlp", "federated", "lending"}. A runnable end-to-end example on a committed synthetic fixture is in the getting-started guide and in notebooks/.

Modules

Plotting (forest plots, reliability diagrams) currently lives in the domain reports. lending's CATE estimation needs the optional fairscope[lending] extra (econml). The federated module audits per-node predictions only — it performs no training and provides no privacy guarantee.

How it differs from AIF360 / Fairlearn

fairscope is complementary to AIF360 and Fairlearn, not a replacement: those toolkits do bias mitigation; fairscope does uncertainty-aware measurement. The table below was verified by inspecting the installed public APIs of AIF360 0.6.1 and Fairlearn 0.14.0 (checked 2026-06; re-confirm if versions change).

* Fairlearn's MetricFrame computes per-subgroup AUC point estimates (e.g. roc_auc_score_group_min), but provides no analytic (DeLong) confidence interval. † AIF360 ships CalibratedEqOddsPostprocessing (calibration-aware equalized-odds postprocessing), not a general per-subgroup temperature/isotonic recalibration interface.

Closest related work — meval (Sutariya & Petersen, 2025, arXiv:2512.17409): a statistical toolbox for stratified, fine-grained model-performance analysis that also provides subgroup metric uncertainty and multiple-comparison corrections (with a medical-imaging focus). fairscope overlaps with it on uncertainty + significance; what fairscope adds is the specific DeLong AUC intervals, the per-subgroup calibration and recalibration interface, the five-axis cross-platform CPFE protocol, and one-call domain audits (healthcare / lending / federated).

Engineering

• Test-driven, with regression tests anchored to authoritative reference values where they exist (DeLong's worked example; statsmodels multiple-testing routines).
• 100% line coverage on the statistical core; CI runs pytest + coverage, ruff, and black across Python 3.9–3.12, and executes the replication notebooks via nbmake.
• Full type hints, NumPy-style docstrings with runnable examples, and explicit input validation (an AUC on a single-class subgroup raises rather than returning a meaningless value).
• Committed fixtures are small, synthetic, and labelled as such; no datasets or trained models are bundled.

Grounded in published research

fairscope ports methods from the author's peer-reviewed and under-review papers; it invents no new mathematics. Each function cites its source. Full venues and identifiers are in CITATION.cff.

• Diabetes risk prediction with external validation + fairness analysis (XGBoost, NHANES→BRFSS) — IEEE CIPHER, 2026.
• A five-axis Cross-Platform Fairness Evaluation for mental-health NLP — under review.
• Privacy-preserving federated learning for diabetes risk across heterogeneous nodes — under review.
• Heterogeneous racial effects in mortgage approval (Causal Forest Double Machine Learning, HMDA) — under review.
• Racial disparities in mortgage lending (RDD / DiD / decomposition, HMDA) — under review.

Citation

If you use fairscope, please cite it via CITATION.cff.

License

MIT © 2026 Rajveer Singh Pall · ORCID 0009-0001-6762-6134