Data preprocessing

Finding and fixing what is **missing, duplicated, inconsistent or wrong** before training — because a model learns the faults along with everything else.

Many systems encode "not recorded" as 0. A blood pressure of 0 is impossible, so the model learns from a value that never occurred.

Absence is **rarely random**. If a value is missing more often for one group, deleting those rows removes that group disproportionately.

It counts **twice**, so that record's characteristics carry double weight in what the model learns.

No. An **error** should be; a **genuine extreme** should not — in fraud detection the outliers are exactly the target.

One **input column** the model is given. Feature selection decides which ones it sees.

A feature containing the answer, or one **not available at prediction time**. Test accuracy looks superb and then collapses in production.

It is unique per row, so the model **memorises** rather than learns — perfect on training data, useless on anything new.

**Irrelevant** (no relationship), **redundant** (duplicates another), **constant** (no information), **identifier** (memorises) — and above all **leaking**.

No. Other features **proxy** for it — postcode, school, employment history — and the model finds those instead.

Describing the same data with **fewer features** while keeping as much of the variation as possible.

More features means an **exponentially larger space**, so data becomes sparse, distances lose meaning, and exponentially more records are needed.

**Selection** keeps a subset of the original columns. **Extraction** builds new features from combinations of them, as PCA does.

Finds the directions along which the data varies **most** and uses them as the new axes, so the first few components carry most of the information.

Some **variation** — it is lossy by construction — and with extraction, **interpretability**: a component cannot be named in plain words.