Hugging Face's logo

Datasets:
drdavidprivacy
/
retainiq-data

Tasks:
Tabular Classification
Modalities:
Tabular
Languages:
English
ArXiv:
Tags:
hr-analytics
fairness
tabular
ai-governance
datasheet
License:
Dataset card Files
xet
 Community
retainiq-data / DATA_REMEDIATION_LOG.md
drdavidprivacy's picture
drdavidprivacy
Upload DATA_REMEDIATION_LOG.md
69fd563 verified
preview code
|
Raw
History Blame Contribute Delete
6.47 kB
metadata 
stage: 5
type: note
status: reference
tags:
  - stage-5
  - note

RetainIQ — Data Remediation Log

The governance artifact closing the data-preparation stage: a complete, auditable record of every data-preparation and remediation decision, with rationale and impact. Modeled on CRISP-DM's Data Cleaning Report (+ Derived Attributes / Generated Records reports) and the Datasheets for Datasets preprocessing section. The point (per CRISP-DM): document, in detail, every decision and action and its possible impact, so a reviewer can verify the preparation didn't distort the data.

Dataset: IBM HR Attrition (synthetic) standing in for RetainIQ training data · Raw copy retained, unmodified: WA_Fn-UseC_-HR-Employee-Attrition.csv · Reproducibility: fixed random_state; all learned steps fit on training only.

Log

ID Stage (CRISP-DM) Issue / reason Action taken Parameters Scope & leakage-safety Applied or Demonstrated Impact
R-01 Select/Clean 3 constant columns carry no information Drop EmployeeCount, StandardHours, Over18 — Whole dataset (fixed, row-wise → no leakage) Applied 35 → 32 cols
R-02 Select/Clean EmployeeNumber is an identifier (memorization/leakage + privacy) Drop it — Whole dataset (fixed) Applied removes ID
R-03 Construct Need a numeric target Define y = Attrition (Yes→1, No→0) — Whole dataset (fixed) Applied 30 features, 1 target
R-04 Format Models need numeric inputs; text categories have no order One-hot encode 7 categoricals (BusinessTravel, Department, EducationField, Gender, JobRole, MaritalStatus, OverTime) drop_first=False Fit on train, applied to all (pipeline) Applied 30 → 51 features
R-05 Format Ordinal scales are ordered categories Keep encoded ordinals as integers (Education, JobLevel, satisfaction/involvement/WLB scales, StockOptionLevel) — do not one-hot — n/a (fixed) Applied preserves order
R-06 Format Features span very different ranges; weight/distance-based models let big-magnitude features dominate Scale numeric features (StandardScaler / z-score) mean & SD per feature Fit on train, applied to all (pipeline) Applied — logistic-regression recipe only (tree/boosting skip scaling) numerics on one 0-centered scale
R-07 Construct (modeling) Class imbalance 5.2:1 → model defaults to "stays" Class weights (class_weight="balanced") leaver weight ≈5.2× (from train class counts) At model fit (train) Applied model attends to minority
R-08 Construct P-01 — proxy/protected-signal bias Compared biased, all-3-attribute mitigation, and age-only CorrelationRemover sensitive considered = Age, Gender, MaritalStatus Fit on train Applied — age-only deployed (v2.0) age-only closes the ADEA gap (EOD 0.369→0.045) with no gender regression / no accuracy loss; see R-16
R-09 Clean Missing values Impute — numeric: median; categorical: mode learned from train Train-only Demonstrated (actual data has 0 missing) technique shown on sample
R-10 Clean Duplicate records Detect & drop exact duplicates — before the split Demonstrated (no dupes in actual data) technique shown on sample
R-11 Clean Inconsistent / invalid entries Standardize (e.g., "Y"→"Yes"); flag out-of-range (Age 199) → impute/correct Age valid range 16–90 — Demonstrated technique shown on sample
R-12 Privacy Re-identification risk (real deployment) k-anonymity via Age generalization (bands) quasi-identifiers: Age, Dept, Gender, JobRole, MaritalStatus — Demonstrated, not applied (synthetic data → no real subjects) uniques shown to drop sharply
R-13 Privacy Individual value disclosure (real deployment) Differential privacy (calibrated noise) illustrative Laplace noise — Demonstrated, not applied per-person scramble; aggregates hold at scale
R-16 Deploy (Stage 5) Which trained model to deploy: biased champion vs bias-mitigated variant Deploy the biased champion v1.0 with the residual disparity disclosed + a monitoring/revisit commitment held-out evidence: age EOD 0.369→0.052 under mitigation, at ROC/PR −0.04 Decision recorded in the Model Card §8 Resolved (deploy-with-disclosed-residual-risk) age/marital-status disparity becomes a named monitoring target

Applied vs. demonstrated (summary)

  • Applied to the model pipeline: R-01–R-08 (drops, target, one-hot, ordinal handling, scaling, class weights, and the two P-01 treatments). The learned transforms — one-hot (R-04), scaling (R-06), CorrelationRemover (R-08) — are fit on training data only; class weights (R-07) are derived from the training class counts. Scaling (R-06) applies only to the logistic-regression recipe; tree and gradient-boosting recipes skip it (scale-invariant).
  • Demonstrated only: R-09–R-13. The actual dataset is clean (0 missing, no duplicates) so no cleaning was needed; privacy PETs aren't applied because the data is synthetic (no real individuals to protect). In a real RetainIQ deployment, R-09–R-13 would be applied. (This demonstrate-vs-apply distinction is itself a documented governance decision.)

Notes

  • Raw data retained unmodified for reproducibility and audit.
  • No leakage: fixed row-wise edits (R-01–R-03, R-05) done on full data; all distribution-learned steps (R-04 one-hot, R-06 scaling, R-08 CorrelationRemover) fit on train only.
  • Bias watch: protected attributes (Age, Gender, MaritalStatus) deliberately retained through prep so fairness can be measured; the use decision is handled via R-08, and the deploy decision via R-16 (resolved above — biased champion deployed with disclosure and a monitoring commitment).

References