PMI CPMAI flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

CRISP-DM phases

Six iterative phases: business understanding, data understanding, data preparation, modeling, evaluation, and deployment. Business understanding frames the problem before any data work begins.

Data drift

A change in input data distribution after deployment relative to training data. It degrades model accuracy even when the underlying relationship between inputs and outcomes is unchanged.

Concept drift

A change in the relationship between inputs and the target outcome over time. The input distribution may look stable while model predictions become progressively wrong.

Data leakage

Use of information during training that would not be available at prediction time. It produces inflated validation scores and disappointing production performance.

Train-test split

Partitioning data into a training set for fitting and a held-out test set for final evaluation. The test set must remain unseen during model tuning.

MLOps

Practices and tooling for reliable machine learning delivery: versioned data, automated training pipelines, deployment, monitoring, and retraining. It extends DevOps to model and data artifacts.

Model acceptance criteria

Pre-agreed thresholds a model must meet to be released, covering accuracy metrics, fairness constraints, latency, and drift tolerance. They are defined before development starts.

Minimum viable model

An early, simplified model built to test feasibility and data sufficiency. It limits sunk cost compared with a long build before any evaluation.

Model retraining trigger

A defined condition that initiates retraining, such as performance below threshold, drift exceeding a limit, or a scheduled interval. Without it, monitoring is incomplete.

Inference deployment modes

Batch scoring runs on schedules over stored data; real-time APIs answer individual requests; edge deployment runs on local devices. Latency, volume, and privacy drive the choice.

Business understanding

The CRISP-DM phase that converts a business problem into measurable objectives, success criteria, and constraints. It precedes data work and anchors later evaluation.

Data quality dimensions

Accuracy, completeness, consistency, timeliness, validity, and uniqueness. Each dimension needs a measurable threshold and a named accountable owner.

Data lineage

A record of data origin, transformations, and movement across systems. It supports error investigation, impact analysis, and audit evidence for AI decisions.

Data owner versus steward

The owner is accountable for a data domain and approves access; the steward executes quality controls and standards day to day. The model owner is accountable for production performance.

EU AI Act risk tiers

Four tiers: unacceptable (prohibited), high-risk (documentation, oversight, logging required), limited-risk (transparency obligations), and minimal-risk. Classification drives compliance obligations.

Human-in-the-loop control

A required human review before an AI output takes effect. It must specify who reviews, what triggers review, and the reviewer's authority to override.

Risk scoring formula

Risk score equals likelihood multiplied by impact. Scores prioritize the AI risk register and set mitigation effort proportionate to exposure.

Anonymization limits

Removing identifiers does not guarantee anonymity, because combining remaining attributes can re-identify individuals. Re-identification testing is required before treating data as anonymous.

Model hallucination

A generative model producing fluent but false or unsupported content. Mitigation includes retrieval grounding, citation requirements, and human verification before use.

Third-party model dependency

Reliance on external model providers introduces risks: service changes, deprecation, pricing shifts, and data handling terms. Contracts should address notice periods and exit paths.

Data minimization

Collecting and retaining only personal data necessary for the stated purpose. It reduces privacy exposure and is a core principle in data protection regimes.

Adoption stages

Awareness, pilot, scaled deployment, and institutionalization. Each stage needs evidence before the next; skipping pilot validation raises failure risk at scale.

Technology acceptance model

Perceived usefulness and perceived ease of use drive intention to use a system. Change plans must address both to move users from awareness to habitual use.

Ethical AI principles

Commonly cited principles: fairness, transparency, accountability, privacy, safety, and human oversight. They are translated into concrete controls rather than statements alone.

Explainability proportionality

The depth of explanation required scales with decision impact. High-stakes outcomes such as credit or hiring demand stronger justification than low-impact internal suggestions.

Role-based AI literacy

Training tailored to role: executives need value and risk framing, practitioners need tooling and data handling, and all staff need acceptable-use rules.

AI governance committee

A body owning AI policy, use-case approval, and exception handling. Effectiveness requires a written charter, decision authority, and a defined review cadence.

Override frequency metric

The rate at which users reject AI recommendations. A high rate signals poor model fit or distrust; a near-zero rate may signal automation bias.

Automation bias

The tendency to accept automated output without sufficient challenge. Mitigations include documented review, sensitivity checks, and requiring justification for accepting recommendations.

AI-assisted estimation

Models that forecast schedule and cost from historical project data. They reduce optimism bias but require challenge of outputs and awareness of historical data limits.

Benefit realization baseline

Metrics captured before AI deployment against which post-deployment value is measured. Without a baseline, claimed benefits cannot be validated.

AI augmentation boundary

A rule for task allocation: AI handles repeatable, data-rich, low-ambiguity work; humans retain novel, high-stakes, and sensitive decisions with accountability.