Feature

A feature is one piece of measurable information about an example — in a spreadsheet, one column.

When you size up a second-hand bike, you check a handful of things: year, kilometres ridden, brand, rust, tyre wear. Each check is a feature. You never inspect "the whole bike" as one blob — you read specific signals and weigh them together. Models work the same way: each example arrives as a list of features, and the model learns how much each one matters for the prediction.

features                                        label
year   km_ridden   brand    rust_level    →    price
2019     8,400     Hero       low         →    ₹34,000
2015    31,000     Honda      high        →    ₹18,500

The thing being predicted is the label; everything used to predict it is a feature. Features can be numbers (kilometres), categories (brand), or things converted into numbers — text becomes token counts or embeddings, images become pixel grids.

For classical models on tables, which features you construct matters more than which algorithm you pick — that craft is feature-engineering. Deep learning's central promise is learning good features automatically from raw input. In both worlds, a feature that will not be available at prediction time must never be used in training: that is data-leakage.

Where to go next#

• Full lesson: What is machine learning?
• Related terms: feature-engineering, label, feature-scaling, embedding