التعويض بالمتوسط أو الوسيط أو المنوال

Instead of inventing random numbers, you fill gaps with a summary value from the column itself. This careful approach is called imputation. 🧮

For numbers, the simplest choice is the mean, the column average. It keeps the overall total roughly intact when gaps are few.

The mean is fragile: a few huge values drag it far from the center. On skewed data, mean imputation can pull every gap toward an unrealistic number.

The median is the middle value, and it shrugs off extreme outliers. For skewed numeric columns it is usually the safer fill.

A quick rule: reach for the median when a column has outliers or a long tail, and the mean only when values are fairly symmetric.

Text and category columns have no average. For them you fill with the mode, the most frequent value, since that is the most likely fit.

A column can tie for most common, so mode() returns a Series of all winners. Pick the first with index 0 when you need one value.

The best practice is to impute each column by its type: median for skewed numbers, mean for symmetric ones, and mode for categories.

For ordered data like time series, carry the last known value forward with ffill, or pull the next one backward with bfill.

Every imputation shrinks the column's natural spread, because filled values cluster at one point. Note this, since it nudges your variance downward.

A pro habit: add a boolean column marking which rows were imputed. That flag lets later analysis know which values were real and which were guessed.

An income column is heavily skewed by a few millionaires. Which imputation fits best?

You now impute with mean, median, or mode by column type, use ffill for ordered data, and flag filled rows so nothing gets hidden.