Exploratory data analysis

Looking before touching

Before any model is trained, the data deserves to be looked at directly. Exploratory data analysis is that look: summary statistics, distributions, correlations between variables, and plain visualisation, all aimed at one question — what does this data actually represent, and where did it come from? Skipping this step and going straight to modeling is how a wrong assumption about the data quietly becomes a wrong assumption baked into the model, discovered only much later when predictions turn out strange.

Univariate, bivariate, and multivariate

Univariate analysis looks at one variable at a time: its distribution (is it roughly symmetric, or heavily skewed to one side?), its typical range, and whether it contains outliers far from the rest of its values. This is the simplest and most foundational pass, and it alone catches a surprising number of data problems — a column of ages containing a value of 200, say.

Bivariate analysis looks at two variables together, most often to check whether they move together — if one rises, does the other tend to rise too, fall, or show no relationship at all? This is where a correlation between two variables first becomes visible, long before any model is trained on them.

Multivariate analysis extends the same question to several variables at once, which starts to reveal interactions a pairwise view cannot — two variables that mean something different in combination than either does alone.

What to watch for

The video moves past definitions into the actual technique, working through univariate, bivariate and multivariate analysis with real distributions and correlations rather than describing them abstractly. That practical grounding is exactly what this lesson needs — the mechanics of looking at data, not just the vocabulary for talking about it.

What a look at the data actually catches

A handful of problems surface reliably at this stage, before they ever reach a model. A skewed distribution — most values clustered low with a long tail of large ones — can make the mean a misleading summary and may call for a transformation before training, a technique covered in feature engineering. An unexpected spike of identical values often signals a placeholder standing in for missing data — a column of ages with suspiciously many entries at exactly zero, say — rather than a genuine measurement, which belongs to the data cleaning lesson immediately following this one. A strong correlation between two input variables can mean they are carrying mostly the same information, which affects how some models, particularly linear ones, interpret their coefficients.

Correlation is a hypothesis, not a conclusion

Finding that two variables move together is often the most exciting moment of an exploratory pass, and it is also where the most common mistake in this whole field happens: treating that association as evidence that one variable causes the other. Two variables can correlate because one genuinely causes the other, because both are driven by some third factor neither of them captures, or by pure coincidence in a small enough sample. Exploratory analysis is where a correlation is noticed; it is not where it gets explained, and the gap between those two things is worth remembering every time a chart looks like it is telling a clean causal story.