Exploratory Data Analysis (EDA) in Python — Beginner Checklist
Read this guide in: English · मराठी · हिंदी
EDA means exploring a table with questions, not staring randomly. Use a pandas checklist: shape, head, info, missing, simple groupby, then write findings. Infosys-style hours example: client A totals 70.
Chala mitrano! Opening a CSV is not EDA. EDA is question → inspect → note. Today a first-pass checklist you can reuse on shop, Instagram-like, Netflix-style or client sheets. Samajla ka?
Quick answer
Vertical first-pass EDA:
- Write one business question.
read_csv→shape→head→columns.info()+isna().sum().describe()for numbers;value_countsfor categories.- One groupby that answers the question.
- Bullet your findings with numbers (then chart in the next guide).
question → inspect → clean notes → groupby → findings
Language: English · मराठी · हिंदी
What do I need before this guide?
- pandas load / filter / groupby / missing basics.
- Course: First-pass checklist.
Before and after (look at the tables first)
Before — Infosys-style client hours dump; no questions.
After — EDA checklist; top client A = 70 hours.
Before: raw hours dump. After: checklist filled; top client A = 70.
What EDA is (and is not)
What it is
- Structured looking guided by a question.
- Finding data-quality issues early.
- Producing short, number-backed findings.
What it is not (on this course)
- Not training sklearn models.
- Not a 40-chart firework show with no sentence.
- Not skipping cleaning forever.
Course tone: Question-driven EDA.
Infosys-style sample
| client | hours |
|---|---|
| A | 40 |
| B | NaN |
| A | 30 |
| C | 20 |
Question: Which client has the most recorded hours?
Vertical pass:
- shape →
(4, 2) - missing → hours has 1 NaN
- Decide fill/drop (here: drop incomplete hour rows or fill — document it)
- If we analyse known hours only: A 70, C 20, B unknown
- Finding: “Client A leads recorded hours at 70 on this extract.”
The checklist (keep it visible)
- Question — one sentence.
- Load — path, sheet, nrows if huge.
- Head / columns — names make sense?
- Grain — one row means what?
- Types — sales numeric? dates parsed?
- Missing — which columns hurt the question?
- Duplicates — unexpected repeats?
- Univariate — describe / value_counts.
- Simple relationships — groupby / crosstab.
- Findings — bullets with numbers.
- Next chart — which visual will tell the story?
Crosstab and correlation caution
- Crosstabs help category × category views.
- Correlation between numbers is not causation.
- Course: Crosstab and correlation caution.
Document findings
Bad finding: “Data looks interesting.”
Good findings:
- “Online shop sales total 190 on the extract (Mumbai 120, Pune 70).”
- “One Netflix-style title has missing watches; filled with 0 for count KPIs.”
- “Client A is 70 hours; B is incomplete in source.”
Course: Document findings.
Shop + Instagram-like quick drills
- Shop: which channel wins? groupby channel → Online 190.
- Engagement: which city wins likes? Mumbai 120 vs Pune 70.
- Same checklist both times — only the question changes.
Timebox tip
- First pass on a small file: aim under an hour.
- Depth comes from sharper questions, not more rainbow charts.
- Stop when you can brief a manager in five bullets.
Ravindra Bagale's Tip
Put the question in a Markdown cell at the top of the notebook. If a cell does not serve that question, move it down or delete it. That is professional EDA. Samajla ka?
Practice task
Using the four-row client hours table:
- Write the question in one sentence.
- Run shape, head, isna().
- groupby client sum on non-missing hours.
- Write three finding bullets including A = 70.
- Name one chart you would draw next (bar of hours by client).
Samajla ka? EDA is question-driven pandas work with written findings — no ML required. Client A 70 is a finding, not a vibe. Aata pudhe jaauya.
Frequently asked questions
What is EDA?
Exploratory data analysis: structured looking before you claim insights or build dashboards.
Is EDA the same as cleaning?
Cleaning fixes types/missing/dupes. EDA uses clean-enough data to answer questions. They overlap in real notebooks.
Do I need seaborn for EDA?
Tables first. Charts help storytelling — covered in the visualisation guides/course chapters.
How long should first-pass EDA take?
For a small CSV, under an hour for a clear checklist. Depth comes from better questions.
Correlation caution?
High correlation is not proof of cause. The course has a caution lesson — read it.
Course links?
EDA chapter: question-driven EDA and first-pass checklist.