How is a data point different from a feature?

The point is the example or unit; a feature is an attribute used to describe it.

Can a data point have no label?

Yes. Unlabeled examples are common in unsupervised and self-supervised learning workflows.

Can two points have identical feature values?

Yes. They may represent different events or entities, and they may even have different outcomes. Identical values do not by themselves prove duplication.

AI Glossary / Governance, ethics and safety / Beginner

Data point

Definition

A data point is one unit of information considered in an analysis or machine-learning task. It may represent an observation, object, event, image, document, or other example. A data point can contain many attributes; it is not necessarily a single number. What counts as one point depends on the question and how the dataset is constructed.

What it means

In a table of deliveries, one row might represent one parcel. Its attributes could include weight, destination zone, dispatch time, and delivery outcome. In another analysis, one data point might represent an entire day's operations, summarized across many parcels. Neither unit is universally correct. The choice changes the meaning of counts, averages, predictions, and errors. It is therefore useful to state the unit of analysis explicitly whenever someone says a model was trained on a certain number of data points.

How it works

A data point is usually represented through features that the model can use and, in supervised learning, a target or label to predict. The same attribute may be a feature in one task and a target in another. A timestamp may define when an observation was made, while an identifier links related observations without necessarily being a suitable predictor. Some points are structured rows; others are images, text sequences, or graphs that require a numerical representation. Data preparation should preserve the distinction between the underlying event or object and its recorded representation. One event can generate several records, and several events can be aggregated into one point.

Three practical examples

In product classification, one photograph may be a data point even though it contains millions of pixel values.
In demand forecasting, one store-day may be a point containing sales, opening hours, and other available features.
In message analysis, one conversation may be the point, or each message may be a separate point. The choice should match whether the model predicts a conversation-level outcome or a message-level property.

Think of it this way

Think of a folder in a filing cabinet. One folder can hold several pieces of information about a case, just as a data point can contain many attributes. The analogy is limited because a point does not have a fixed natural size: an analyst might create separate folders for each event or combine them into a summary folder. The important part is consistency about what each unit represents and how related units are connected.

Worked example

Suppose a table contains twelve orders from three customers. Customer A placed ten orders and two were late. Customers B and C each placed one order, and both were late. At the order level, there are twelve data points and four late outcomes, so the late-order rate is 4/12 = 33.3%. At the customer level, the individual late-order rates are 20%, 100%, and 100%. Their unweighted average is 220% divided by three, approximately 73.3%. These answer different questions: the first weights each order equally, while the second weights each customer equally. Neither calculation should be presented without its unit. If the same customers appear in both training and test data, the evaluation may also answer a different question from performance on entirely new customers.

Illustrative case study

Fictional case: Ash Service reports a large training dataset for predicting equipment faults. A review finds that many rows are repeated readings from the same small set of machines. The team adds machine identifiers and collection periods to its documentation, then evaluates separately on later readings and on machines excluded from training. It still uses the repeated measurements where appropriate, but stops describing the row count as the number of independent machines observed. The case illustrates how defining a data point clarifies both coverage and the limits of evaluation.

Common misconceptions and mistakes

A data point is not always a dot on a two-dimensional chart; many points require hundreds or thousands of features. A row count is also not automatically a count of independent observations. Repeated measures, duplicated records, and related documents can create dependence. Missing attributes do not make a record cease to be a data point, but they affect how it can be used. Finally, labels are not necessarily perfect ground truth. They may be measurements, judgments, or operational records with their own uncertainty and errors.

Business use

Before collecting or purchasing data, define the unit, relevant attributes, target, and collection time. Ask whether the dataset represents unique entities, repeated events, or transformed versions of earlier records. Document how duplicates, missing values, and corrections are handled. Choose train-test splits that match the intended deployment question, such as future events or unseen entities. When information concerns people, consider the sensitivity of the combined attributes rather than assuming a single row is harmless. Clear definitions make data quality checks, access decisions, and performance reports easier to interpret.

Check your understanding

A dataset has 5,000 daily readings from fifty machines. Does it necessarily contain 5,000 independent examples of different machines?

Answer: No. It contains 5,000 recorded machine-day observations, with repeated observations from fifty machines. The dependence and intended prediction task should guide analysis and evaluation.

Related terms

Instance
Feature
Label
Sample size
observation
dataset
unit of analysis
repeated measures
data provenance

Learn next

Take a familiar spreadsheet and complete the sentence “Each row represents one ...”. Then identify which columns are features, targets, identifiers, and metadata for a specific task. Finally, list ways that two rows could be related, because those relationships may affect sampling and evaluation.

Sources

Alexander Jung, A Dictionary of Machine Learning, data point and dataset entries: https://aaltodictionaryofml.github.io/ADictML_Main.pdf Lee and colleagues, Deduplicating Training Data Makes Language Models Better, for the importance of repeated examples and train-test overlap: https://arxiv.org/abs/2107.06499

Previous: Data pipeline Next: Data poisoning