Data Management and Governance
Concepts a candidate must know
This domain covers how data is sourced, described, protected, and kept fit for AI use. Understand data quality dimensions - accuracy, completeness, consistency, timeliness, uniqueness, and validity - and how each affects model behavior. Know the difference between structured, semi-structured, and unstructured data, and between metadata, master data, and reference data. Be able to explain data lineage, provenance, and cataloging, and why they matter for reproducibility and audit. Understand data protection concepts: classification tiers, access control, encryption, anonymization versus pseudonymization, and retention. Know the roles of data owner, steward, and custodian, and the principle of data minimization.
Traps that appear in questions
A frequent trap is treating anonymization and pseudonymization as equivalent; pseudonymized data can often be re-identified via linkage, so it usually remains personal data. Another is assuming more data always improves a model - biased, duplicated, or stale data can degrade it. Questions may conflate data governance with IT security, though governance is broader (quality, ownership, policy, lifecycle). Watch for scenarios where lineage is missing and reproducibility is claimed anyway, and for retention policies that conflict with minimization.
Worked example / rule of thumb
Rule of thumb: if you cannot trace a training record back to its source and transformation history, treat the model as non-auditable. Example: replacing names with customer IDs is pseudonymization, not anonymization, because the mapping table allows re-identification; true anonymization must make re-identification reasonably impossible.