---
title: Feature
description: 'An input variable or derived representation that a machine-learning model actually uses to produce a prediction.'
canonical_url: 'https://darkfactory.dev/glossary/feature'
markdown_url: 'https://darkfactory.dev/glossary/feature.md'
collection: glossary
date_published: '2026-08-03T00:00:00-04:00'
date_modified: '2026-08-26T00:00:00-04:00'
---

# Feature


## Definition

A feature is an input variable or derived representation that a machine-learning model actually uses to produce a prediction. It can come directly from a recorded field, such as account age, or from a transformation, such as transactions per week. Deep models may learn internal features from rawer inputs such as tokens, pixels, or audio samples.

The feature definition includes how and when the value is computed. "Customer balance" could mean the current ledger value, the value at application time, a rolling average, or a post-decision value. Those variants may share a name while carrying different information and leakage risk.

## Feature quality

A feature must be available at the moment the prediction is made. Training-serving skew occurs when training uses a value or transformation that production cannot reproduce. Target leakage occurs when the feature contains information created after, or because of, the outcome the model is supposed to predict.

Missingness can itself carry information, but only if the collection process is understood. A default value that silently combines "zero" with "unknown" can create a feature the model cannot interpret consistently.

## Distinguish it from nearby terms

- A **raw field** is collected data. A feature is the value presented to the model after any selection or transformation.
- A **label** is the target or reference answer used for training or evaluation.
- A **parameter** is learned during training. A feature is supplied for an example at training or inference time.
- An **embedding** is a learned vector representation that can serve as one feature or a set of features.

## Operational significance

Maintain lineage from source field through transformation to model input. Version the computation, test parity between training and serving, monitor missingness and drift, and review whether a feature is permitted for the decision. Removing a protected attribute does not remove its proxies.

## Check your understanding

A churn model uses "days since cancellation" and scores extremely well. The feature is known only after the event, so it leaks the label and cannot support a real pre-cancellation prediction.

## Related terms

- [Label](https://darkfactory.dev/glossary/label)
- [Dataset](https://darkfactory.dev/glossary/dataset)

## Evidence and further reading

- [NIST AI Resource Center Glossary](https://airc.nist.gov/glossary/)
- [Google Machine Learning Glossary](https://developers.google.com/machine-learning/glossary/)