fairscope¶
Subgroup-stratified, calibration-aware fairness auditing for machine-learning models — grounded in peer-reviewed methods.
fairscope packages statistical machinery that mainstream fairness toolkits do not provide
as first-class, subgroup-stratified functions:
- DeLong confidence intervals for per-subgroup AUC
- Expected Calibration Error per subgroup, with reliability diagrams
- Bootstrap significance testing of subgroup performance differences, with multiple-comparison correction
- a subgroup-stratified interface to standard recalibration (temperature scaling, isotonic regression)
- a novel five-axis Cross-Platform Fairness Evaluation (CPFE) protocol
- per-class fairness metrics (symmetric disparate impact, equalized odds difference)
It ports established methods rather than inventing new ones; only the CPFE protocol is
presented as novel, and every function cites its source. fairscope is for rigorous,
uncertainty-aware measurement — it is complementary to mitigation-focused toolkits such as
AIF360 and Fairlearn, not a replacement.
Install¶
pip install fairscope
Heavy NLP dependencies (torch, transformers, captum) install via pip install fairscope[nlp];
SHAP via pip install fairscope[shap]. The base install stays light.
Where to go next¶
- Getting started — a real, runnable audit in a few lines.
- Healthcare audit — the one-call clinical fairness report.
- CPFE protocol — the five-axis cross-platform evaluation.
- API reference — every public function.
Source and releases: github.com/Rajveer-code/fairscope · MIT license.