Glosarium
Contents

Data Engineering › Transformation & Analytics SQL

Exploratory Data Analysis

Getting to know a dataset before modelling it.

Data Analyst · Junior · core

Also known as: exploratory data analysis, EDA, data exploration

Exploratory data analysis (EDA) is the first date with a dataset: its size and shape, what is missing, what looks wrong, what might be interesting — before any model or dashboard commits to an interpretation. An hour of EDA routinely saves a week of modelling the wrong thing, because most datasets arrive with surprises their documentation never mentions.

shape → distributions → missingness → oddities → relationships → questions
(rows? cols? types?) (histograms) (where? how much?) (spikes? negatives?) (what moves together?)

Work in that order. Confirm grain and size first (is each row what you think?), then distributions (profiling), then missingness and odd values, and only then relationships. Questions that survive this pass deserve modelling; most early hypotheses do not survive contact with the histograms.

The classic mistakes:

  • Modelling before looking. A pipeline on unexamined data bakes in every surprise — leakage, duplicates, shifted definitions — as silent features. Look first, always.
  • Exploring without writing anything down. EDA that lives in a scrolled-away notebook is lost. Note data quirks, dead ends and open questions; they become the data docs and the model caveats.
  • Treating exploration as confirmation. EDA finds hypotheses; it does not test them. A pattern spotted in exploration needs fresh data (or a test) before it becomes a claim — otherwise it is circular.
  • Ignoring missing data and outliers. The weird rows are often the story (a broken sensor, a fraud pattern) or the threat to it. Either way they deserve attention before the pretty charts.
  • Plots without reading them. Generating twenty histograms and glancing at none is worse than useless — it creates the feeling of diligence. Look at every plot you make, and ask what would be surprising.

The output of EDA is not a deliverable but readiness: a mental model of the data’s quirks, a list of questions worth modelling, and the judgement to distrust the dataset in the right places. See data visualisation for showing what you found.