EDA Workflow
From improvisation to a repeatable method
A repeatable EDA sequence — goals → audit → profile → relations → feature candidate list → memo — turns exploration from improvisation into a disciplined, shareable process.
What you will learn
- Structure EDA into stages with exits
- Produce a feature-candidate list with reasons
- Write the hand-off memo modeling consumes
What is this?
A workflow gives EDA a beginning, middle and exit: (1) define the question and success metric, (2) audit the data, (3) profile columns and relationships, (4) draft feature ideas and risks, (5) write a one-page memo. It converts 'look at the data' (unbounded) into a scoped artifact that demonstrably finishes and hands off cleanly to modeling.
Why do you need it?
- Unstructured EDA meanders — 'interesting patterns' all day, no decisions.
- A scoped checklist survives hand-offs: anyone can re-run it on new data.
- The memo is deliverable: it captures decisions future-you and reviewers rely on.
- A workflow separates 'analysis' from 'analysis paralysis'.
Mental model
A chef's mise-en-place and tasting notes before the kitchen rush
EDA workflow
Prep is structured, notes are taken, and the pass (the hand-off) is a defined moment. No chef stares at a fridge 'exploring' at 7pm — stages, then the service.
How it works
- Stage 0 — Question: what decision does this dataset serve?
- Stage 1 — Audit: shape, dtypes, nulls, dups, describe.
- Stage 2 — Profile: univariate shapes; then pairs (corr, crosstabs).
- Stage 3 — Ideate: list candidate features + risks (leakage, sparsity, units).
- Stage 4 — Memo: one page: findings, decisions, open questions, next steps.
The workflow as a pipeline.
In production
- A data team's onboarding guide literally lists the five stages, so juniors hit a solid 70% on day one.
- An experiment postmortem reruns the audit+profile stages on the winning cohort to avoid surprises.
- A modeling hand-off gate requires the memo — no memo, no model training quota.
Best practices
- Time-box exploration: the question changes at stage boundaries, not mid-flow.
- Record every decision with a why — the memo is the audit trail.
- Re-run the workflow when the dataset contract changes (new source, schema).
- Use a standard template so memos are comparable across analyses.
Common mistakes
- Skipping the question stage, exploring aimlessly for a day.
- Ending with analysis and no written decisions (tribal knowledge).
- Duplicate workflows per analyst, so findings aren't comparable.
Debugging tips
- Stuck? Re-read the decision from stage 0 and re-scope the exploration.
- If a memo says 'everything looks fine' with no risks, the workflow wasn't rigorous.
- Version the memo alongside the dataset — findings drift when data shifts.
Interview checkpoint
Quick quiz
The FIRST stage of a disciplined EDA workflow is:
The EDA memo is a delivery artifact, not an internal flourish.