Explaining the Past

Data science explains the past through data exploration. Before you predict anything, you describe what already happened: what the typical case looks like, how much values vary, which variables move together, and where the anomalies hide. This is the descriptive half of data mining.

Description is not a warm-up act for modeling — many projects deliver their entire value here. A well-built summary of last year’s failures, churned customers, or fraudulent transactions is often the first time an organization has actually seen its own data.

Description versus prediction

The two halves of data mining answer different questions and are judged by different standards. Descriptive mining summarizes the data you have; its claims must be faithful to that data. Predictive mining — predicting the future — generalizes to data you have not seen; its claims must survive contact with new observations, which is why it needs models and held-out evaluation.

Descriptive work decomposes naturally along two axes:

• Univariate analysis — one variable at a time: distributions of categorical and numerical variables.
• Bivariate analysis — pairs of variables: categorical vs. categorical, numerical vs. numerical, and categorical vs. numerical.

A small worked example

Seven weeks of transaction counts from a shop’s logs:

Week	1	2	3	4	5	6	7
Transactions	10	12	14	16	18	20	22

The univariate summary already tells a story. The center:

$\bar{x} = \frac{1}{n}\sum_{i=1}^{n} x_i = \frac{112}{7} = 16, \qquad \text{median} = 16$

and the spread around it:

$s^2 = \frac{1}{n}\sum_{i=1}^{n}(x_i - \bar{x})^2 = \frac{112}{7} = 16, \qquad s = 4$

Mean equals median, so the series is symmetric; the steady climb from 10 to 22 is a trend, not noise. Two sentences of arithmetic have explained the past — and raised the predictive question a regression model would answer next: what should week 8 be?

In practice

Descriptive analysis today lives in pandas (df.describe(), value_counts()), Polars, and SQL aggregations, with visualization in matplotlib, seaborn, or Plotly. Summary statistics and plots are also the first diagnostic for model problems: most “model bugs” are data issues that five minutes of exploration would have exposed. Modern dashboards automate the descriptive half of data mining entirely — what they cannot automate is deciding which description matters.

Common pitfalls

• Reporting only means when the distribution is skewed or multimodal — the median and the histogram tell the truth.
• Confusing an observed association with a cause; description establishes what, not why.
• Exploring after modeling instead of before, so obvious data errors contaminate the model.
• Dismissing outliers as errors before checking whether they are the most informative rows in the data.

Summary

Explaining the past is descriptive data mining: univariate and bivariate exploration that summarizes what happened. It is valuable in its own right, and it is the indispensable foundation for the predictive work of modeling.