简体中文
EnglishWhat Is CRISP-ML(Q) and How Does It Relate to ML Model Governance?
Updated
CRISP-ML(Q) is the standard process model for machine learning development presented on ml-ops.org, and it relates to ML model governance by giving governance work a defined lifecycle to attach to: phases for understanding the business problem, preparing data, building and evaluating models, and deploying and monitoring them. The site treats it as the reference process model for ML development and covers governance as a separate topic area, so the practical connection is that governance processes—review, documentation, monitoring, and lifecycle control—map onto the stages CRISP-ML(Q) defines. This explanation is for readers deciding how to structure an ML project and where to place governance checkpoints; it describes the framework as the site presents it rather than prescribing a specific compliance regime.
What CRISP-ML(Q) is on ml-ops.org
The site lists CRISP-ML(Q) under its core learning sections with the description: "You will learn about the standard process model for machine learning development." That is the extent of the detail in the available page evidence—the site positions it as the standard process model, not as a proprietary method.
The "(Q)" signals a quality-oriented extension of the classic CRISP-DM style process, but the available material does not enumerate its phases or quality gates. Treat the phase list above as the general shape of a standard ML process model, and consult the site's CRISP-ML(Q) section for the authoritative phase definitions before you build templates on top of it.
Why a process model matters for ML projects
The site's framing explains the need. It describes MLOps as an emerging field gaining momentum among data scientists, ML engineers, and AI enthusiasts, and it cites the Continuous Delivery Foundation SIG MLOps in differentiating ML model management from traditional software engineering. Two of the listed MLOps capabilities are directly relevant here:
- MLOps enables automated testing of machine learning artifacts, including data validation, ML model testing, and ML model integration testing.
- MLOps enables supporting machine learning models and datasets to build these models as first-class citizens within CI/CD systems.
A process model is what tells you when those tests and validations belong. Without a defined lifecycle, data validation and model testing become ad hoc, and the site's stated goal of reducing technical debt across ML models is harder to achieve.
How CRISP-ML(Q) connects to ML model governance
The site covers ML Model Governance as its own section, described as presenting "an overview of governance processes." The relationship works in both directions:
| Governance concern | Where CRISP-ML(Q) provides the hook |
|---|---|
| Business problem and requirements understood before design | The site stresses this as one of the most important phases in any software project and says it applies equally to ML-based software |
| Data quality and validation | Data preparation and validation stages, reinforced by the MLOps capability for automated data validation |
| Model evaluation and testing | Model building and evaluation stages, reinforced by ML model testing and integration testing |
| Production monitoring and change management | Deployment and monitoring stages, reinforced by the site's coverage of model serving patterns, deployment strategies, and monitoring practices |
| Accountability across the lifecycle | Governance processes documented per phase, rather than as a separate afterthought |
The site also notes that MLOps "must be a language-, framework-, platform-, and infrastructure-agnostic practice." That matters for governance design: because CRISP-ML(Q) is a process model rather than a tool specification, you can implement its stages on whatever stack you already run, and governance controls can be expressed as process requirements instead of platform features.
A concrete example
Suppose a team is preparing a credit-risk scoring model. Using CRISP-ML(Q) as the spine, the business-understanding phase produces a written problem statement and success criteria; the data phase produces a validation report; the modeling phase produces evaluation results against held-out data; and the deployment phase produces a monitoring plan. Governance then has four natural checkpoints—one per phase—where a reviewer signs off on the artifact the phase produced. The alternative, bolting governance on after deployment, tends to produce documentation that describes the model but not the decisions behind it.
Where to go next on the site
The site's structure supports following this thread in order:
- Motivation for MLOps — scenarios of change that need managing and the iterative nature of ML-based software development.
- Designing ML-powered Software — understanding the business problem and requirements before designing.
- End-to-End ML Workflow Lifecycle — a high-level overview of a typical ML workflow.
- Three Levels of ML-based Software — data, ML models, and code, including data engineering pipelines and ML pipelines.
- MLOps Principles — practices to test, deploy, manage, and monitor models in production.
- CRISP-ML(Q) — the standard process model itself.
- MLOps Stack Canvas — an application- and industry-neutral framework for specifying architecture and infrastructure.
- ML Model Governance — governance processes for models in production.
If your goal is governance specifically, read CRISP-ML(Q) and ML Model Governance together; the first supplies the stages, the second supplies the controls. If your goal is architecture, the MLOps Stack Canvas is the more direct entry point.