Feature

Starter
How Machines Learn
Technical readers
Programme & data teams

Also called: input variable, attribute, features.

Detail

In five seconds

What it is

A feature is a piece of input information a model uses to help produce an output, such as a prediction, score or grouping.

In plain words

A feature is one piece of information you give a model so it can produce an output.

An everyday example

A technician at a water station in Rome checks a pump each morning using a clipboard. She writes down pressure, motor temperature, vibration level and whether the sound is steady or irregular. If several readings look unusual together, she marks the pump for inspection. Each reading on the clipboard gives part of the picture; no single reading always tells the full story.

Where the story stops working

A human technician can use wider context, judgement and physical knowledge that a machine learning model may not have. Also, the clipboard analogy makes features look hand-chosen and simple, but in practice features can be automatically derived, transformed or encoded from complex data such as text, images or time series.

Remember thisA feature is one piece of input information used by a model. It may be measured directly or derived from other data, but it is not the model’s target output.

Understand it

Why it matters

If you hear that a model used certain features, you can ask practical questions: What information went in? Was anything important missing? Were some inputs poor quality, biased or out of date? Did the system use a proxy variable that may unfairly stand in for something sensitive? These questions affect accuracy, fairness, privacy, explainability and whether the output is suitable for real decisions. For non-specialists, understanding features helps you scrutinise a system without needing to build one yourself.

From raw case details to features to output

Show that features are the input pieces of information for a model, and distinguish them from labels and internal parameters.

01
One case
Start with one example, such as one pump reading or one survey response record.
Draw one card on the left labelled 'Example'. Inside, list 4 short fields such as pressure 82, temperature 67, vibration 0.4, hour 03:00.
02
Features extracted
Each usable input field becomes a feature. Some are direct measurements; some may be transformed versions.
From the example card, draw arrows to 4 small boxes labelled 'pressure', 'temperature', 'vibration', 'hour'. Use a note saying 'features = input variables'.
03
Model uses features
The model receives the feature values together, not one by one in isolation.
Place a central rounded rectangle labelled 'model'. Draw all feature boxes into it as parallel arrows.
04
Output produced
The model returns an output such as a class, score or estimated number.
Draw one arrow from the model to a right-side box labelled 'output: unusual reading risk 0.86'.
05
What features are not
A label is the target used in training, and parameters are internal learned settings, not input features.
Below the main flow, add a dashed training-only lane: 'known label: inspected fault / no fault' beside the example, with a comparison arrow to the output. Inside the model, show faint small dots labelled 'parameters learned during training'.

A left-to-right flow diagram. On the left is one case with several raw details. In the centre, those details are shown as named feature boxes entering a model. On the right is the output. A separate note below shows that, during training only, a label sits beside the example and is used to compare with the output, while parameters stay inside the model.

Hover or tap each box to reveal a short note: feature, label, output or parameter. A toggle switches between training view and inference view by showing or hiding the dashed label lane.

Describe this without the diagram

The diagram reads from left to right. First, there is one example card containing raw details: pressure 82, temperature 67, vibration 0.4 and hour 03:00. Next, each detail becomes a feature box labelled pressure, temperature, vibration and hour. These four feature boxes all point into a central box labelled model. From the model, one arrow points to an output box labelled unusual reading risk 0.86. Below this main path, a separate dashed lane appears only for training. It shows a known label, such as inspected fault or no fault, placed beside the example and compared with the model output. Inside the model, small faint dots are labelled parameters learned during training. The message is that features are inputs from the data, labels are target answers used during training, and parameters are internal settings inside the model.

With reduced motion: A static left-to-right diagram with five labelled stages: example, feature boxes, model, output, and a dashed training-only label lane below. No animation is used; the training and inference difference is shown with visible labels and line styles.

In machine learning, features are the pieces of information that go into a model. If a system is trying to flag an unusual water pump reading, features might include pressure, temperature, vibration level and time of day. The model uses these inputs to produce an output, such as “normal” or “check this pump”.

Several related words are used here. “Feature” is the most common machine learning term. “Input variable” usually means the same thing, especially when people want to stress that it is one column or field going into a model. “Attribute” is also close, but it is more common in some older materials, in database discussions, and in some forms of unsupervised learning. The words overlap, but they are not always used in exactly the same community.

A feature is not the same as a label. A label is the target output a model is trained to predict, such as “fraud” or “not fraud”. A feature is evidence the model can use, such as transaction amount or country code. A feature is also not the same as a parameter. Parameters are internal numerical settings learned by the model during training; features come from the data.

Features can be directly observed, like age or rainfall, or derived from other data, like average spending in the last 30 days. During training, the model looks for patterns connecting features to labels or to structure in the data. During inference, it receives feature values for a new case and produces an output. Good features often matter as much as model choice, because the model can only work with the information it is given.

How “Checking a water pump with a clipboard” maps onto the real thing

In the story	In the AI system
Pressure, temperature, vibration and sound notes on the clipboard	Features or input variables
The technician’s inspection mark	The model’s output, such as a prediction or flag
The pattern across several readings together	How a model uses multiple features jointly
A single morning check for one pump	One example given to the model

A worked example

Before the donor reporting package enters clearance, the evaluation office uses a simple classifier to identify progress reports that may require closer review.

Starting from: One report was submitted on time, has an incomplete results field, includes an explanation for variance and has an unresolved action from the previous review.

The four report attributes are selected as features.
The categorical values are encoded numerically, for example yes as 1 and no as 0, so the model can process them.
The model compares this feature vector with patterns learned from previously reviewed reports and assigns a follow-up category.
The humanitarian affairs officer checks the source report and records the reason for any follow-up.

Result: The report is placed in the review queue before clearance. The model helps prioritise the file, but does not itself establish poor performance or determine a funding decision.

In a United Nations workplace

A humanitarian affairs officer on temporary assignment to an evaluation function is preparing a donor reporting package for quarter-end clearance. A model sorts implementing partner progress reports for follow-up using features such as timely submission, missing results fields, unexplained variance and unresolved actions from the previous review.

Each listed attribute is a feature: an input the model uses to assign a follow-up category. Some are recorded directly, while others are derived from the report. The category is the output, not another feature. Feature choice determines what evidence the model notices and what it overlooks.

Before you try thisProgress reports may contain personal, partner-confidential or protection-sensitive information. Do not upload them to any tool not approved for that material, as this could expose information outside authorised controls. Use minimised data in an approved environment, check for misleading proxy features, and require human review before any performance or funding decision.

Where you may hear it

During a vendor demo for an evaluation office
At a data-preparation meeting before a monitoring report is cleared
In an ICT governance board review of a proposed prediction tool
While agreeing fields for an implementing partner reporting template

Avoid confusion

Often confused with

Common misconceptions

Common misconception: A feature is just any column in a spreadsheet.
Correction: Often a feature does come from a column, but not every column should be used. Some fields may be identifiers, notes, duplicates, or protected information that should not feed the model. Also, one feature can be created from several columns, such as a 30-day average.
Common misconception: Features and labels are basically the same because both are data fields.
Correction: They play different roles. Features are the inputs used to make an output. A label is the target answer used during training in supervised learning. Mixing them up can cause major errors, including leakage, where the model gets access to information it would not have in real use.
Common misconception: If you have more features, the model will always perform better.
Correction: More inputs can help, but they can also add noise, cost, bias and instability. Irrelevant or low-quality features may reduce performance, make the system harder to explain, or create unfair proxies for sensitive traits.
Common misconception: A feature is the same as a parameter.
Correction: A feature comes from the data for each case, such as age or temperature. A parameter is an internal numerical setting the model learns during training. Features change from case to case; learned parameters usually stay fixed during inference unless the model is retrained.

Quick check

Which option describes a feature most accurately?

Go deeper

The simple version

A feature is one piece of information you give a model so it can produce an output.

The technical version

In machine learning, a feature is an input variable representing one measurable or derived property of an example. Features may be numeric, categorical, binary, text-derived, image-derived or time-based, though many models require them to be encoded into numerical form before use. In supervised learning, the model is trained to map feature vectors to labels or target values. In unsupervised learning, it uses features to find structure such as clusters or lower-dimensional representations without target labels. The term overlaps with input variable and, in some contexts, attribute, but usage varies by field. Features should be distinguished from labels, which are target outputs, and from parameters, which are internal learned values of the model.

Reviewed external resources

EncyclopaediaFeature (machine learning) (opens in a new tab)Wikipedia
This Wikipedia page gives a broad overview of the term and related language, which suits readers who want a quick reference and are comfortable checking details critically.
Intermediatelink checked 2026-07-31
CourseMachine Learning Crash Course (opens in a new tab)Google
This course resource is useful for beginners who want practical explanations of machine learning inputs and model building in a structured format.
Intermediatelink checked 2026-07-31
VideoStatQuest (opens in a new tab)StatQuest with Josh Starmer
This video source is helpful for learners who prefer step-by-step spoken explanations and visual teaching of core machine learning ideas.
Starterlink checked 2026-07-31
VideoVideo explanations: Feature (opens in a new tab)YouTube search
This YouTube search results page helps readers find multiple video explanations of features, which suits people comparing teaching styles rather than looking for one verified source.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Feature (opens in a new tab)Reddit search
This Reddit search results page lets readers see how practitioners informally discuss features and common confusions, which may help curious readers but should not be treated as authoritative.
search results, not one pagelink checked 2026-07-31

Where this came from

Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.

The one thing to remember

A feature is one piece of input information used by a model. It may be measured directly or derived from other data, but it is not the model’s target output.