drdavidprivacy/retainiq-data

Datasheet — RetainIQ Training Data (IBM HR Attrition)

This is the dataset card for RetainIQ's training data. Created for The AI Governance Lab course following the Datasheets for Datasets template (Gebru et al., 2018), because no official datasheet exists for this dataset. Authoring one ourselves is the point: most real datasets arrive undocumented, and writing the datasheet is how a governance-minded team takes responsibility for what it's about to build on. Answers marked "not documented" reflect genuine gaps in the original source — and those gaps are themselves findings.

ℹ️ On the data file & licensing. The IBM HR Analytics Employee Attrition sample is freely public and universally used for teaching, but its exact license terms are loosely specified. To stay clean, we do not re-host the raw file here — we link to the original source (see Distribution & License below). This card documents the data; download it from the canonical page.

At a glance

Dataset name	IBM HR Analytics Employee Attrition & Performance
Role in this course	Training data for RetainIQ (employee-attrition prediction)
Size	1,470 rows × 35 columns — one row per (fictional) employee
Prediction target	Attrition (Yes = 237 / No = 1,233; 16.1% positive — imbalanced)
Nature	Synthetic — fictional data created by IBM data scientists
Missing values	None
Protected attributes present	Age, Gender, MaritalStatus
License	Open / educational; exact terms loosely specified (see note above)
Source	Linked to the original (not re-hosted) — see Distribution & License

1. Motivation

• Why was it created? As fictional sample data to demonstrate IBM Watson Analytics' HR-attrition capabilities — not collected from real HR operations and not created for our use.
• Who created / funded it? IBM data scientists. Details beyond that are not documented.
• Key provenance fact: it showcases analytics on invented people, so its patterns are designed, not observed.

2. Composition

• Each row = one fictional employee. 35 columns, grouped:
  • Target: Attrition (Yes/No).
  • Identifier (not a feature): EmployeeNumber (unique per row; non-contiguous).
  • Constant / no information: EmployeeCount (=1), StandardHours (=80), Over18 (='Y').
  • Demographics / protected: Age (18–60), Gender (Female/Male), MaritalStatus (Single/Married/Divorced).
  • Education: Education (1–5, ordinal), EducationField (6 categories).
  • Role & organization: Department (3), JobRole (9), JobLevel (1–5), BusinessTravel (3), OverTime (Yes/No).
  • Compensation: MonthlyIncome (1,009–19,999), MonthlyRate, DailyRate, HourlyRate, PercentSalaryHike (11–25), StockOptionLevel (0–3).
  • Tenure & history: TotalWorkingYears (0–40), YearsAtCompany, YearsInCurrentRole, YearsSinceLastPromotion, YearsWithCurrManager, NumCompaniesWorked (0–9), TrainingTimesLastYear (0–6).
  • Engagement / satisfaction (1–4 ordinal scales): JobSatisfaction, EnvironmentSatisfaction, JobInvolvement, RelationshipSatisfaction, WorkLifeBalance; plus PerformanceRating (only 3 or 4 ever appear).
  • Other: DistanceFromHome (1–29).
• Important nuance: many "numeric" columns are encoded ordinal categories (the 1–4 / 1–5 scales), not true quantities — they must be understood as categories.
• Data-quality notes: no missing values; PerformanceRating has no low ratings (only 3–4); the suspicious tidiness is consistent with synthetic origin.
• Sensitive data: contains protected attributes. Because the data is synthetic there is no real PII — but in a real RetainIQ deployment these would be real employees' personal data, with all the consent and privacy obligations that implies.
• Labels: the Attrition label ships with the data; how it was assigned is not documented.

3. Collection process

• How was it collected? Not applicable in the usual sense — the data was generated by IBM, not collected from real employees. The generation method is not documented.
• Timeframe / sampling / ethical review: not documented (and largely moot for synthetic data).
• Governance note: the emptiness of this section is the lesson. For a real RetainIQ, this is where consent, data sources, collection timeframe, and review would be recorded — and a reviewer would expect them.

4. Preprocessing / cleaning / labeling

• As distributed: already clean — no missing values, a mix of numeric and categorical fields.
• Done for RetainIQ: see the companion DATA_REMEDIATION_LOG.md in this repo for the complete, auditable record of every preparation decision (drops, encoding, scaling, class weights, bias treatments, and the demonstrated cleaning/privacy techniques).
• Raw data retained? Yes — this canonical copy is kept unmodified.

5. Uses

• Intended (this course): teaching end-to-end ML and AI governance through RetainIQ attrition prediction (binary classification), including fairness analysis (the protected attributes make this possible).
• Cautions / not suitable for: it is synthetic — patterns are designed, not real — so nothing learned here should be taken as a real-world claim about what drives attrition. The clean, tidy nature can breed false confidence. Treat the encoded scales and the PerformanceRating quirk with care.

6. Distribution & License

• Availability: public on Kaggle and many mirrors.
• Our choice: because the exact license terms are loosely specified, we link to the canonical source rather than re-hosting the file in this repo. Download the data here: https://www.kaggle.com/datasets/pavansubhasht/ibm-hr-analytics-attrition-dataset
• For reproducibility: the course keeps a verified local copy; learners pull the same file from the link above.

7. Maintenance

• Maintained by: no active maintainer; it is a static dataset.
• This copy: versioned for the course; could be checksummed for integrity.

Connection to RetainIQ

We use this dataset as a stand-in for RetainIQ's training data — one employer's workforce. Teaching simplification (recorded honestly): a real RetainIQ would train on consented, real, possibly multi-customer workforce data; we flag the gap rather than pretend the dataset is something it isn't.

References

• Gebru, T., Morgenstern, J., Vecchione, B., Vaughan, J. W., Wallach, H., Daumé III, H., & Crawford, K. (2018). Datasheets for Datasets. https://arxiv.org/abs/1803.09010
• Dataset page (de facto documentation): https://www.kaggle.com/datasets/pavansubhasht/ibm-hr-analytics-attrition-dataset