Data science approaches (blueprint)

Purpose: Deeper, project-agnostic guides for data science methodologies and process frameworks. Each approach describes its structure, phases, and mapping to the lifecycle.

Guide · Updated · Source

Audience: Teams adopting Data science & machine learning; project-specific ML decisions stay in docs/architecture/ and docs/adr/.

Methodology selection

Pick a process scaffold first (goals, data reality, deployment), then layer operations (MLOps) when models must live in production under change control. CRISP-DM fits most greenfield analytics and ML projects; MLOps is not a replacement for CRISP-DM but the engineering wrapper around deployment and monitoring. Experiment management and A/B testing address how you learn during development and how you prove impact in production — they complement, rather than substitute for, a phase model. If you are unsure where to start, read CRISP-DM: Cross-Industry Standard Process for Data Mining for phase definitions, then MLOps: Machine Learning Operations when the path to production is real rather than hypothetical.

Core knowledge: Data science & machine learning body of knowledge — ML lifecycle, statistics, evaluation, MLOps, responsible AI.

Bridge: Data Science ↔ SDLC ↔ PDLC bridge — how data science maps to delivery and product lifecycles.

Data science methodology selection

How process scaffolds, operational layers, and learning loops combine when teams choose data science approaches.

  1. Methodology selectionTeams pick a process scaffold before layering operations and measurement.
  2. Lane A ──► handoff ──► shared outcomeThe primary lane delivers governed process through to shared delivery outcomes.
  3. Lane B ──► inspect / adapt ──► feedbackA parallel lane inspects signals and adapts methodology choices from evidence.
  4. Note: illustrativeTiming overlaps in practice while dependency order stays fixed.

The diagram is illustrative: in practice, experiment tracking starts early (often during Data Preparation / Modeling in CRISP-DM terms), and A/B tests may run only after a first deployment — but the dependencies (sound process → disciplined experiments → reliable production → measured impact) stay the same.

Approach Focus When to use Guide
CRISP-DM Six-phase cross-industry standard process for data mining and ML projects Default starting point for most ML projects; industry-proven; tool-agnostic CRISP-DM: Cross-Industry Standard Process for Data Mining
MLOps Operationalizing ML — CI/CD for models, automated retraining, model monitoring When ML models run in production and require ongoing management MLOps: Machine Learning Operations
Experiment management Systematic approach to running, tracking, and analyzing ML experiments During model development; comparing approaches; hyperparameter optimization DATA-SCIENCE.md §3
A/B testing Controlled online experimentation to measure model impact on product metrics When validating model impact in production; PDLC P5 outcome measurement DATA-SCIENCE.md §3

Related techniques: For encoding, feature stores, and evaluation metrics used inside these phases, see Feature engineering & feature stores and Model evaluation, selection & validation.

Keep project-specific model documentation in docs/product/ and experiment logs in docs/development/, not in this file.