Exploratory Data Analysis

The fastest way to reach a wrong conclusion is to model data you've never looked at. A column that's secretly text, a metric that's half missing, a distribution with a long tail — none of these announce themselves. They quietly bias every test and model built on top of them, unless you look first.

Exploratory Data Analysis is that first look. It isn't a single technique but a family of quick, structured checks — describe each column, count the categories, and see what correlates — that turn a raw table into something you understand before you commit to a method.

The Three Questions of Exploration

1. 1What is each column? Its type, how complete it is, and whether it holds what you think it does.
2. 2How is each value spread? The center, the spread, the shape — symmetric or skewed, one peak or several.
3. 3What moves with what? Which variables rise and fall together, and where a relationship is worth a closer look.

Answer these and the shape of the whole dataset comes into focus — long before a single hypothesis is tested.

Describe: Center, Spread, and Shape

Descriptive statistics summarise a numeric column in three numbers' worth of meaning, as the histogram above shows — where it centers, how far it spreads, and what shape it takes.

When the mean and median drift apart, the column is skewed and the median is the honest summary. The descriptive statistics guide covers when to trust each.

Count: Frequency Analysis

Numbers have distributions; categories have frequencies. Frequency analysis counts how often each category appears — which reveals dominant classes, rare levels that will trouble a model, and the imbalance that quietly distorts accuracy.

• Dominant categories that a model may simply memorize
• Rare levels that break one-hot encoding or thin out a group
• Class imbalance — the reason accuracy can look great and mean nothing

Relate: Correlation Analysis

Once each column is understood on its own, exploration asks how they move together. Correlation measures the strength and direction of a linear relationship, flagging pairs worth a deeper look.

Where EDA Leads

Exploration isn't an end in itself — it sets up every later decision. What you find routes you to the right next step.

Exploration in the SKARI Statistical Lab

SKARI treats exploration as the first analysis family — the starting point before comparison, relationship, or predictive work. The Data Editor and Statistical Lab profile and summarise before you commit to a method.

• Descriptive statistics per column — center, spread, shape, and outlier flags
• Frequency analysis for categorical columns and class balance
• Correlation analysis with coefficients and significance
• Column profiling in the Data Editor to catch type and completeness issues first

Frequently Asked Questions

Is EDA just making charts?

Charts help, but EDA is the structured habit of describing, counting, and relating every variable — so you understand the data, not just picture it.

How much EDA is enough?

Enough to answer the three questions — what each column is, how it's spread, and what correlates — and to catch anything that would invalidate your planned method.

When do I stop exploring and start testing?

Once the data holds no surprises that would break your chosen analysis. EDA that never ends is procrastination; EDA that's skipped is a trap.

Key Takeaways

Exploration turns a raw table into understanding. Describe each column, count the categories, see what correlates — and the right test, transform, or model is already half-decided before you run anything.