Feature

A feature is an input attribute or representation used by a model. It may be a measured value, a category, a derived count, or a learned representation such as an embedding. A feature definition describes how to obtain it; a feature value is the result for a particular example.

Define “recent purchases” precisely

Suppose a recommendation model uses a customer's purchase count during the previous seven days. The definition must identify the prediction time, which order statuses count, how cancellations are handled, and the interval boundaries. A value of 4 means little without those choices.

For a prediction at noon, tomorrow's purchase must not enter that historical training example. This is information leakage: using information unavailable at the intended prediction time. Google's glossary provides the basic distinction between model features and target labels.

Keep preparation and use consistent

If training counted paid orders but the live service counts all created orders, the feature name may match while the meaning differs. Align definitions, units, missing-value handling, and transformation versions across training and serving.

A feature prepared overnight can be cheap to retrieve but may omit this morning's activity. That tradeoff needs a freshness requirement. A feature store can help manage feature data, but it does not decide the correct business definition or eliminate leakage by its name alone.

Reference: Google: Machine learning glossary.