license: mit
library_name: sklearn
pipeline_tag: tabular-classification
tags:
- tabular-classification
- hr-analytics
- fairness
- logistic-regression
- ai-governance
- model-card
datasets:
- REPLACE-WITH-YOUR-ORG/retainiq-data
Model Card — RetainIQ (v2.0, bias-mitigated)
Employee-attrition risk model · Hireloom, Inc. · Structured per Mitchell et al. (2019), "Model Cards for Model Reporting." Worked example for The AI Governance Lab. Educational use only — not for real employment decisions.
1. Model Details
| Model name | RetainIQ — Employee Attrition Risk Model |
| Owner | Hireloom, Inc. (provider and deployer — self-oversight / dogfooding) |
| Version / date | v2.0 · June 2026 (supersedes v1.0 biased champion; see §8) |
| Model type | Binary classification — logistic regression (C=0.1, class_weight=balanced) |
| Bias mitigation | Fairlearn CorrelationRemover on Age only (the ADEA-protected attribute) |
| Pipeline | One-Hot + CorrelationRemover(Age) + StandardScaler → LogReg; 50 features; scikit-learn 1.7.2; seed 42 |
| Artifact | RetainIQ_Champion_AgeOnly.pkl (8,969 B) + readable coefficients (_Inside.csv) |
| Integrity (SHA-256) | 32a1a21bf9b4984eb5466f955bbef9a729a5535ca81ea279fb48a23513d74057 |
| Contact | [Hireloom AI Governance lead — placeholder] |
2. Intended Use
- Primary use: flag employees at elevated attrition risk so HR can prioritize retention outreach. Decision-SUPPORT only; monthly batch-scored, capacity-ranked list.
- Intended users: Hireloom HR / People team and people-managers.
- Out of scope: NOT for termination, discipline, compensation denial, or any adverse action; not autonomous; not validated outside Hireloom's workforce.
3. Factors
Protected groups assessed: Age (ADEA-protected 40 line), Gender, Marital Status. Outputs a probability converted to a flag via a chosen threshold.
4. Metrics
ROC-AUC and PR-AUC (threshold-independent), plus recall, precision, F1 at 0.50. Because leavers are ~16%, PR-AUC and recall are prioritized over plain accuracy. The production operating threshold is a separate, capacity-driven choice set by HR.
5. Evaluation Data
Sealed test set: 294 employees (20% stratified hold-out, seed 42), 47 real leavers (16.0%). Quarantined from Module 5; opened once for the final grade.
6. Training Data
1,176 employees (80% stratified split), 16.2% leavers. Protected attributes retained through preprocessing so fairness could be measured; the deployed pipeline then applies age-only CorrelationRemover (removes the Age signal; retains Gender/Marital Status).
7. Quantitative Analyses
Sealed-test performance of the deployed v2.0 (age-only) champion, with the biased v1.0 for comparison:
| Metric | Deployed v2.0 (age-only) | Biased v1.0 (rejected) |
|---|---|---|
| ROC-AUC | 0.811 | 0.812 |
| PR-AUC | 0.534 | 0.588 |
| Recall | 0.702 | 0.681 |
| Precision | 0.407 | 0.381 |
| F1 | 0.516 | 0.489 |
Held-out fairness (equalized-odds difference; lower = fairer): Age 40+ 0.045 (was 0.369 — gap closed; 40+ recall 0.44→0.69); Marital Status 0.281 (improved); Gender 0.026 (held fair). Caveat: PR-AUC is marginally lower (0.534), so the gain is clearest at the chosen operating point; per-group counts are small (40+ n=16, Divorced n=4).
Why age-only: a three-way bake-off showed all-attribute CorrelationRemover worsened gender (EOD 0.033→0.137) and cost more accuracy, while age-only closed the ADEA gap with no gender regression and no accuracy cost.
8. Ethical Considerations
- Deployment decision (R-16) — RESOLVED: deploy the age-only bias-mitigated champion (v2.0), prioritizing the ADEA-protected age group (largest gap + greatest legal exposure). The biased v1.0 and the all-3-attribute variant are documented as considered-and-rejected alternatives.
- Residual / monitoring targets: the marital-status spread persists, and gender balance must be watched to confirm the untargeted groups don't drift — both are named monitoring targets (Module 9).
- Self-oversight / dogfooding: Hireloom is both provider and deployer; independent fairness review recommended.
- Misuse risk: flags are decision-support, not verdicts; adverse use would be inappropriate and likely unlawful.
9. Caveats & Recommendations
- Synthetic data stand-in: revalidate on real Hireloom data before any deployment.
- Threshold: set/document the production operating point by HR capacity; re-run the fairness audit AT that threshold per protected group.
- Monitoring: watch drift, the gender EOD, and the marital spread; define retrain/retire triggers.
- Legal/compliance (not legal advice): obtain review of ADEA, Title VII, NYC Local Law 144, EEOC guidance, EU AI Act (high-risk employment) and GDPR special-category obligations before real deployment.
Reference
Mitchell, M., et al. (2019). Model Cards for Model Reporting. FAT* '19. arXiv:1810.03993.