Data Science Foundations and Methodology
Data Science Foundations and Methodology
This domain tests the project lifecycle rather than a single library: how a business question becomes a data requirement, an analytical approach, a model, an evaluation, and a deployed decision. Expect questions on the IBM Data Science Methodology, CRISP-DM, supervised versus unsupervised problems, and the roles of data engineers, analysts, and scientists.
Core rules
- IBM's methodology has ten stages: business understanding, analytical approach, data requirements, data collection, data understanding, data preparation, modeling, evaluation, deployment, feedback.
- CRISP-DM condenses this into six iterative phases: business understanding, data understanding, data preparation, modeling, evaluation, deployment.
- The analytical approach follows the question: classification when labels exist, regression for continuous outcomes, clustering for segmentation without labels, association rules for co-occurrence.
- Structured data lives in tables; unstructured data includes text, images, and audio. Data understanding profiles data; data preparation fixes it.
Worked example
A subscription service asks why cancellations rose. The exam-ready chain is: the deliverable is a retention decision, not a model. Define churn operationally as no renewal within 30 days. Choose classification because historical churn labels exist. Specify features: tenure, plan, support tickets, usage decline. After cleaning, train and evaluate with recall emphasized if missing a likely churner is costly. Deploy weekly scoring and monitor drift and campaign lift in the feedback stage.
Traps
- Starting with an algorithm before defining the decision, target, and success metric.
- Treating data understanding and data preparation as the same stage.
- Presenting CRISP-DM as a strict waterfall; iteration is expected.
- Stopping at model accuracy and omitting deployment, monitoring, and feedback.