Does a label have to be a class name?

No. It can be a number, vector, sequence, or structured annotation, depending on the task.

Can a labeled data point have an incorrect label?

Yes. Label availability and label reliability are separate questions.

Can the same example support different tasks?

Yes. Different choices of inputs and targets can create different learning tasks from the same underlying record.

AI Glossary / Mathematics and optimization / Beginner

Labeled data point

Definition

A labeled data point is an example for which a target value has been supplied or determined. In supervised learning, it is often represented as a pair (x,y), where x contains the features and y is the label. Having a label means the target is available; it does not guarantee that the target is correct.

What it means

The label answers the particular question defined by the learning task. The same photograph might carry a category label, an object count, or a collection of bounding boxes. A sales record might carry the later demand that a forecasting model is supposed to predict.
Labels can come from people, instruments, administrative records, or rules that construct a target from available data. A person's judgment, a sensor reading, and a delayed business outcome can each be useful, but each has different possible errors.
The observation unit must be clear: a customer, customer-month, and transaction are different examples. Attaching a target at the wrong level changes its meaning.

How it works

Define the input, target, and observation unit before collecting labels. Specify how the target is determined, when it becomes available, and which feature values would be available at prediction time. Then associate the target with the correct example using a stable identifier or other reliable correspondence.
Review ambiguous cases against the annotation policy. If the target remains unknown, retain that uncertainty instead of inventing a value. For supervised training, the learning algorithm compares its prediction for x with the supplied y through a loss function.
The pair is therefore task-dependent. Information can legitimately serve as a target for one task and as an input for another, provided the resulting task and timing make sense. A label's presence alone does not establish that the example is suitable for evaluating future performance.

Three practical examples

A document with an agreed topic category is a labeled example for topic classification.
A building's recorded characteristics paired with a measured energy total form a labeled example for an appropriately defined energy-prediction task.
A product photograph paired with reviewed object boundaries forms a labeled example for object detection, where the target is structured rather than a single number.

Think of it this way

Think of a practice exercise accompanied by an answer key. The exercise gives the learner something to work from, and the key provides a target for comparison. The analogy has limits: real labels can be disputed or noisy, and many learning tasks predict uncertain outcomes rather than questions with one timeless correct answer.

Worked example

Consider an original prediction task for equipment inspections. Each example represents one machine on one inspection date. The features are its age and vibration reading at that inspection. The label is whether a qualifying fault is recorded within the following seven days: 1 for a fault, 0 for no fault.
Machine A is inspected on day 1 with features (4 years, 2.1 units). A fault is recorded on day 4, so its target becomes 1. Machine B is inspected on day 1 with features (2 years, 1.3 units), and complete observation through day 8 finds no qualifying fault, so its target is 0.
Machine C is inspected on day 6, but the dataset ends on day 8. Its seven-day outcome is not yet established. Assigning 0 merely because no fault has been recorded by day 8 would confuse an incomplete observation window with a confirmed negative. Under this simple policy, C remains unlabeled until the window is resolved.
Suppose predictions for the two completed examples are A:1 and B:1. One of the two predictions matches its target, giving accuracy 1/2=50% on these two examples. Including C as a made-up negative could change that calculation without adding valid outcome evidence. The example illustrates target construction and timing; two cases are far too few to estimate deployment performance reliably.

Illustrative case study

In a fictional service company, reviewers label incoming messages by the action a support team should take. They discover that some messages contain several requests, while the initial policy requires one action. Rather than allowing reviewers to improvise different conventions, the team clarifies how to choose the primary action and records unresolved cases.
It preserves the annotation-policy version. When the policy later permits multiple actions, the team treats those labels as a changed target definition. A labeled row needs context, not just a nonempty field.

Common misconceptions and mistakes

Do not equate labeled with verified, objective, or free from disagreement. Do not assume every label was created manually. Do not treat an unobserved event as a confirmed negative when the observation window is incomplete.
Avoid including information in the features that would only be available after the prediction is supposed to occur. Also distinguish a training label from a prediction generated by another model: a model-produced target may be useful under an explicit method, but its errors and provenance remain relevant. More labeled examples do not automatically compensate for systematically misleading target definitions.

Business use

A labeled dataset needs an agreed target policy and reliable example-target links. Teams should document the observation unit, target source, timing, uncertain cases, and any review process. Sampling complete examples for inspection can reveal mistakes that summary counts miss.
For planning, distinguish the cost of collecting inputs from the cost of obtaining dependable targets. Some outcomes arrive naturally but only after a delay. Other targets require specialist judgment or repeated review. This distinction helps explain why a large store of raw data may still provide relatively few suitable labeled examples for a particular task.

Check your understanding

A record has no observed fault yet, but only two days of a seven-day outcome window have passed. Is its seven-day label necessarily zero?

Answer: No. The outcome window is incomplete, so the negative target has not been established under the stated policy.

Related terms

Learn next

Next, describe one labeled example from your own setting, including what was known at prediction time and when the target became available. Then examine label noise and dataset splitting to understand why valid individual examples still require careful evaluation design.

Sources

Primary terminology reference: Alexander Jung and contributors, The Aalto Dictionary of Machine Learning, labeled data point and label entries: https://aaltodictionaryofml.github.io/ADictML_Main.pdf . The inspection-window example and annotation workflow are original illustrations.

Previous: Labeled data Next: Labeled example