NeuroNomixer
  • Home
  • Blog
  • Visual Guides
  • Authors
  • Contact
Sign InSign Up
HomeBlogAuthorsContactPrivacy Policy

© 2026 NeuroNomixer — Built with Next.js & Tailwind CSS

Visual Guides/Interrogating a Dataset
Data & Analysis

Interrogating a Dataset

A fresh dataset never confesses on its own. You have 120 sensor readings from a warehouse and four problems hiding in them. Work the standard questioning sequence until all four are on the board.

Sentinel values
Duplicated batch
Impossible category
Skewed column

Sign in to save progress

EDA is a loop, not a checklist

Exploratory data analysis has a rhythm: ask a question, look at the answer, refine the question, look again. The five panels below are the standard opening questions (shape, types, missingness, distributions, duplicates, relationships), but they are not a form to fill in once. A discovery in one panel changes what every other panel shows, so you circle back. And you write each finding down the moment you see it, because a finding you did not record is a finding you will rediscover at 2 a.m. in production. The board on the right is your notebook.

How this playground stays honest: the 120-row dataset is generated once with a seeded generator (seed 941) and is identical on every visit. Four issues were planted in it deliberately. Every count, histogram, skewness, and correlation on this page is computed live from those 120 rows. Nothing is faked.

Step 1: What are you holding?

Before any chart, ask the boring questions: how many rows, how many columns, what type is each column, and how many distinct values does it take? The deployment manifest says this warehouse has exactly four zones: north, south, east, and west. Inspect each column and compare what the data says against what the manifest promises. If a categorical column has more distinct values than it should, one of them is lying.

Rows

120

Columns

6

Unique reading_id

112

columndtypenon-nullunique
reading_idstring120 / 120112
zonestring120 / 1205
temp_cnumber116 / 12056
humidity_pctnumber120 / 12080
voc_ppbnumber120 / 120108
battery_vnumber120 / 12038

Findings Board

0 / 4
  • Sentinel values (not found yet)

    A column with zero nulls but an impossible minimum. Step 2 has a switch.

  • Duplicated batch (not found yet)

    120 rows but fewer unique ids. Step 4 has a scan.

  • Impossible category (not found yet)

    Compare zone's unique count against the manifest in step 1.

  • Skewed column (not found yet)

    One histogram in step 3 hides its shape until you change its axis.

Real analysts keep exactly this: a running log of what the data admitted and which question extracted it.

The loop discipline

Notice what happened when you flagged the sentinel: the missingness map changed, the humidity histogram snapped into shape, and every correlation involving humidity became meaningful. One upstream answer rewrote three downstream ones. That is why EDA cannot be a checklist you run top to bottom once: each finding invalidates conclusions you already drew, so you loop back and re-look. In practice the sequence is: shape and types first (cheapest, catches schema lies), then missingness (before trusting any statistic), then per-column distributions, then duplicates, then relationships. Then again, until a full pass turns up nothing new.

The same discipline applies to what you do next. The sentinel readings and the duplicated batch should be fixed at ingestion, not papered over in analysis code. The mislabeled zone needs a human decision: is "Sout" a typo of south, or a fifth zone nobody documented? And the skewed VOC column is not a defect at all, just a column that wants a log scale. Interrogation tells you which of those four conversations to have.

← All GuidesNext Guide →