AI Glossary / AI and machine learning / Beginner

Feature

Definition

A feature is one measurable input field a model learns from, such as a customer's age, an order's total or the hour a call arrived.

What it means

Features are the columns of the model's world. Everything a model is allowed to know about each case arrives as features: numbers, categories, dates, text, anything measurable, translated into values.

The model can only be as good as its features. An algorithm cannot invent information it was never shown, so the work of choosing and constructing features, feature engineering, often matters more than the choice of algorithm.

Features differ in kind: continuous ones such as revenue, discrete ones such as contract tier, dense ones filled on nearly every record and sparse ones mostly empty. Each kind carries its own handling rules.

Not every field deserves to be a feature. Redundant, leaky or irrelevant fields add noise, cost and sometimes cheating, such as a field that secretly encodes the answer.

For a business, features are where domain knowledge enters the model: the decision that change in order frequency matters more than order frequency itself is a business judgment wearing a technical name.

How it works

List candidate fields from the available data, then ask of each: is it knowable at prediction time, is it lawful and ethical to use, does it carry signal.

Transform raw fields into model-ready features: encode categories, extract parts of dates, compute deltas and ratios that capture the business meaning.

Check each feature's quality: coverage, stability, distribution and correlation with the outcome and with other features.

Select or reduce: drop fields that add nothing, combine those that duplicate, and document what was kept and why.

Monitor features in production: drift in an input's distribution is the earliest warning that a model's world has changed.

Give every feature an owner and a definition in the data dictionary: the person accountable when the field's meaning or pipeline changes, because undocumented features decay silently.

Three practical examples

• Churn prediction: tenure, monthly spend, support contacts and change in usage as features.
• Fraud detection: transaction amount, merchant category, time of day and distance from home.
• Demand forecasting: price, promotions, seasonality fields and weather.

Think of it this way

Think of describing houses to someone pricing them sight unseen. You cannot hand them the house; you hand them a list: bedrooms, area, year built, distance to the station. That list is the features, and the pricing can only be as good as the list.

The limit: leave garden condition off the list and every valuation ignores gardens. Feature choice is silent about what it excludes, and the model shares the silence.

Worked example

adding one feature that mattered

Suppose a lender's default model uses forty fields and plateaus. The figures are illustrative.

A loan officer mentions that applicants who ask about early repayment behave differently, so the team adds one feature: whether the applicant asked.

Validation accuracy barely moves, but losses on the riskiest band fall 6 per cent, because the new feature sharpens exactly the decisions the old ones fudged.

The lesson the team records: the best features are often knowledge that staff had and the data never said out loud.

Illustrative case study

A fictional retailer's model mysteriously scored 99 per cent accuracy at predicting which orders would be refunded, until an audit found a feature named days_until_refund_processed, populated after the refund happened.

The model was not predicting refunds; it was reading their paperwork. Removing the leaky field dropped accuracy to an honest 81 per cent.

The invented lesson: a feature that knows the future is not a feature, it is a leak. Check knowability before celebrating accuracy.

The invented follow-up: the retailer added a leak checklist to model review, asking of every feature when it becomes knowable, and the question has caught two further leaks since, both in fields nobody had thought to suspect.

Common misconceptions and mistakes

• More features mean a better model. Relevant features do; irrelevant and redundant ones dilute, slow and sometimes mislead.
• Feature engineering is obsolete in deep learning. Deep models learn representations but still depend on what inputs they are given; the work moved, it did not vanish.
• Features are purely technical. Their selection encodes business and ethical judgment, including what the model is forbidden to consider.

Business use

Review the feature list in every model proposal as a business document: what the model may know is a policy decision.

Ask of every suspiciously accurate model: which feature could be leaking the answer.

Fund feature monitoring: drift reports on inputs are cheaper than post-mortems on outputs.

Invite front-line staff into feature reviews: they hold the candidate features that data teams cannot see, and one field from the floor is often worth ten from the warehouse.

People also ask

What is feature engineering?

The work of choosing and constructing the input fields a model learns from, including transformations like deltas, ratios and encodings.

What is a leaky feature?

One that contains information unavailable at prediction time, often a disguised trace of the outcome itself.

How many features does a model need?

As many as carry independent signal; beyond that, each must justify its keep against held-out performance.

Check your understanding

Why does the quality of a model's features usually matter more than the choice of algorithm?

Answer: Because algorithms can only combine the information they are given; a field that captures the real driver contributes more than any amount of algorithmic sophistication applied without it.

Related terms

• Feature engineering
• Encoding (for Features)
• Dense feature
• Discrete feature
• Delta feature
• Data leakage
• encoding

Learn next

AI for Business Owners explains how to read a model's feature list as a statement about what the model is allowed to know: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

More in AI and machine learning

• Extremal optimization - Extremal optimization is a heuristic search method that preferentially changes poorly performing components of a candidate solution.
• Factor graph - A factor graph represents how a function of several variables can be written as a product of smaller functions called factors.
• Farthest-first traversal - Farthest-first traversal orders points by repeatedly selecting the point whose distance to its nearest previously selected point is greatest.
• Feature cross - A feature cross is a new input made by combining two or more existing fields, so the model can learn patterns that only appear when the fields interact.
• Feature cutoff - A feature cutoff is a threshold that drops feature values below a set frequency or importance, keeping the input space small enough to be useful.
• Feature extraction - Feature extraction is deriving informative measurements from raw, unstructured material, such as turning a photo into edges and textures or a document into...

Sources

[S2] Google, Machine Learning Glossary; [S12] Aalto University, Dictionary of Machine Learning; [S14] KNIME, Data Science Glossary; [S3] NIST AI Resource Center glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: FCO-IM Next: Feature cross