In supervised learning, what makes an example 'labeled'?
Choose an answer, then check it.
Data Science & AI Literacy · Foundations
Supervised Learning
On this page 9 sections
In 30 seconds
Supervised learning is machine learning that learns from labeled examples, cases that come with their known answers attached. A model studies many labeled examples, finds the pattern connecting the input to the answer, and then predicts answers for new, unlabeled cases. It powers two classic tasks: classification The supervised task of sorting cases into categories, such as spam versus not-spam. Full entry →, sorting things into categories such as spam or not-spam, and regression The supervised task of predicting a numeric value, such as a house's selling price. Full entry →, predicting a number such as a house's price.
Why this matters
Supervised learning runs behind everyday tools: email filters that keep spam out of your inbox, banks checking loan applications for risk, platforms flagging fraud, and systems that support medical decisions. It matters academically because it is the clearest example of how machines learn from data, and later topics, features, testing, overfitting, and bias, build on it. It matters practically because knowing what supervised learning can and cannot do helps you judge claims about AI, from 'this model was trained on labeled data' to 'this model sorted people into categories.' You do not need math to grasp the core idea: learn from answers, then predict.
The college version
Learning from labeled examples
Supervised learning is machine learning that trains on examples that come with their answers attached. Each labeled example A case in a dataset that carries its known correct answer, such as an email marked spam or not-spam. Full entry → pairs an input with its known correct output: a photo of an apple together with the word 'apple,' an email together with the verdict 'spam,' a house listing together with the price it actually sold for. The name comes from the role of the human teacher: someone has gone through the data and supplied the correct answer for each case, the way a tutor marks up a workbook. The algorithm's job is to study those example pairs, find the pattern that connects the input to the answer, and then apply that pattern to cases it has never seen. Everything else about supervised learning, the tasks, the loop, the quality worries, follows from this one idea: learn from examples that carry known answers.
The two tasks: classification and regression
Supervised tasks come in two main kinds. Classification sorts cases into categories. The Maple Street community garden volunteers share an inbox, and a filter trained on past messages sorts each new email as either 'junk' or 'not junk' before a volunteer reads it. Two categories, a clear verdict for each message: that is classification, and the spam-versus-not-spam filter is the standard example of it. Regression predicts a number instead of a category. A real-estate agent in Fairview wants to estimate what a three-bedroom house with a new roof and a finished basement will sell for. A regression model trained on past sales learns how features like size, condition, and neighborhood relate to the price, then produces a dollar figure for a new listing. If the answer is a category, it is classification; if the answer is a number, it is regression.
The supervised learning loop
The whole method is a three-step loop. First, gather labeled examples: past emails marked spam or not-spam, past house sales with their prices. Second, show the model those examples so it can learn the pattern, the relationships between the features of each case and its label The known answer attached to an example, the value a supervised model is trained to predict. Full entry →. The model does not memorize the examples; it finds the general rule that connects inputs to answers. Third, put the model to work on new cases that have no labels and let it predict. When the predictions are wrong, more or better labeled examples can be added and the model retrained. Microsoft Learn describes exactly this cycle: train on data that includes both features and known labels, evaluate the predictions, and refine. The loop is simple to describe and powerful because the same recipe works for spam filters, price estimates, and medical risk flags alike.
Features and labels
Two terms carry the whole machinery. The features are the inputs, the facts a case comes with: the sender and subject of an email, the square footage and age of a house. The label is the answer: 'spam' or 'not-spam,' a sold price. Supervised learning exists precisely because there are cases whose features are known but whose labels are not, and the model's job is to fill in the missing labels. How features are chosen, measured, and turned into numbers is a rich subject of its own, and features and labels get their own lesson in this course. Here it is enough to hold onto the division: features in, label out.
Where supervised learning shows up
Supervised learning is not an exotic lab technique; it runs behind everyday tools. Email filters use classification to keep spam out of inboxes, a use Google Cloud's explainer treats as the classic example. Banks and financial firms use supervised models to assess risk, for instance predicting whether a loan applicant is likely to default. In healthcare, supervised models support diagnosis by flagging patients who may be at risk of a condition such as diabetes, based on clinical measurements, so a clinician can review, a support role, not a replacement for medical judgment. And price prediction The model's output for a new case whose label is unknown, produced after training. Full entry →, from houses to airline tickets, is regression in action. In each case the shape is the same: labeled history in, predictions out.
Good labels matter
The label is the ground truth The verified, correct answer used as a label, the standard a model's learning is measured against. Full entry → the model learns from, and labels come from human effort. Someone must read each email and decide, measure each house, or review each scan, and that work takes time and money. IBM's explainer is blunt about it: labeling data takes a lot of time and effort. It follows that label quality is the ceiling on model quality. If the labels are wrong, the model learns the wrong pattern: a spam filter trained on emails mislabeled by an automated script will confidently repeat the script's mistakes. Bad labels mean bad learning, no matter how clever the algorithm is. When you hear about a model failing in the wild, the first question worth asking is not 'what algorithm?' but 'what was it taught?'
Eli explains
The same idea, in plain words
Explain it like I’m 10
Supervised learning is learning from examples that come with the answers already written on them. Imagine a stack of flashcards, each with a picture on the front and the name written on the back. A teacher shows the stack to a learner; the learner studies front and back together, and after enough cards it can name a brand-new picture it has never seen. That is the whole recipe: examples with answers, a pattern learned, predictions on new cases. If the answer is a category, like junk or not junk, it is classification; if the answer is a number, like a price, it is regression. The trickiest part is that someone has to write the answers on the cards, and if those answers are wrong, the learner learns the wrong thing.
Picture it like this
Think of a student teacher learning to grade essays from a stack of already-graded ones. Every essay in the stack has a grade written on it by an experienced teacher. The student reads the graded essays, notices that clear structure and few spelling errors tend to earn high marks, and then grades a fresh essay that has no mark on it yet. The old grades are the labels, the essays' qualities are the features, and the fresh essay is the new case. Supervised learning does exactly this, except the 'student' is a program and the grading happens thousands of times over.
Where the picture stops working
The grading analogy breaks down because a program never 'understands' essays the way a human student does; it only finds statistical patterns in the examples. It also hides the cost: a real teacher grades each essay once, but a supervised model needs many thousands of labeled examples, and someone had to create every one of those labels by hand.
Worked example
Greenway Bike Shop keeps a year of sales records: for each bike sold, the model, the frame size, the buyer's distance from the shop, and the price paid. The owner wants a model that predicts what a new bike should sell for. This is a regression task: the answer is a number. The features are the model, frame size, and buyer distance; the label is the price. The shop feeds the model hundreds of past sales, labeled examples, and the model learns how each feature relates to price: carbon frames sell for more, farther-away buyers accept higher prices, and so on. When a new bike arrives with no price yet, the model reads its features and outputs an estimate, say $1,240. The estimate is a starting point, not a promise: prices change with seasons and stock, and the model only knows what the past sales taught it.
Key takeaway
Supervised learning trains on labeled examples, cases with known answers, learns the pattern connecting features to labels, and predicts answers for new cases. Classification sorts into categories, regression predicts numbers, and because labels come from human effort, bad labels produce bad learning.
Quick check
3 questions here, of 5 in this lesson’s practice set. Answers stay hidden until you check.
A photography club trains a model on pictures of birds, each picture labeled with the species name, so it can name birds in new photos. What kind of supervised task is this?
Choose an answer, then check it.
After a supervised model has learned from labeled examples, what does it do with a brand-new case that has no label yet?
Choose an answer, then check it.
Study tools & related lessonsYou’ll learn to · Common mistakes · Easily confused · Key vocabulary · Related
You’ll learn to
- Define supervised learning as machine learning from labeled examples that carry known answers.
- Distinguish classification, sorting cases into categories, from regression, predicting a numeric value.
- Explain the supervised learning loop: study labeled examples, learn the pattern, predict labels for new cases.
- Identify real-world uses of supervised learning, including email filtering, medical diagnosis support, and price prediction.
- Explain why label quality matters, since labels come from human effort and models learn whatever the labels teach them.
Common mistakes
Thinking classification and regression are interchangeable.
Classification returns a category, spam or not-spam. Regression returns a number, a price or a salary. Ask what the answer looks like.
Believing the model creates its own labels from nowhere.
Every label was made by someone, a human read, measured, or verified it. Labels are labor, and the model inherits both the work and the mistakes.
Expecting a model trained on messy labels to fix itself.
Bad labels teach bad patterns. If the training labels are wrong, the predictions will be wrong in the same way, no matter which algorithm is used.
Assuming the model memorizes the examples instead of learning a pattern.
A supervised model generalizes: it finds the pattern and applies it to new cases. Memorization without a pattern would be useless for new emails or new houses.
Easily confused
Classification vs. Regression
Classification sorts each case into a category, such as spam or not-spam; regression predicts a numeric value, such as a selling price.
Labeled examples vs. Unlabeled data
A labeled example carries its known answer, which is what supervised learning trains on; unlabeled data carries no answers and belongs to unsupervised learning, a separate topic.
Feature vs. Label
A feature is an input fact about a case, such as square footage; the label is the answer the model is trained to predict, such as the sold price. Features and labels get their own lesson.
Spam filter vs. Price predictor
Both are supervised, but the filter outputs a category, classification, while the price predictor outputs a number, regression.
Key vocabulary
- labeled example
- A case in a dataset that carries its known correct answer, such as an email marked spam or not-spam.
- label
- The known answer attached to an example, the value a supervised model is trained to predict.
- feature
- A measurable input fact about a case, such as a house's square footage or an email's sender; features and labels have their own lesson.
- classification
- The supervised task of sorting cases into categories, such as spam versus not-spam.
- regression
- The supervised task of predicting a numeric value, such as a house's selling price.
- training
- The step in which a model studies labeled examples and learns the pattern that connects features to labels.
- prediction
- The model's output for a new case whose label is unknown, produced after training.
- ground truth
- The verified, correct answer used as a label, the standard a model's learning is measured against.
Sources & references
- What is supervised learning? — Google Cloud
- What is supervised learning? — IBM
- Introduction to Machine Learning Concepts (Microsoft Learn training module: fundamentals-machine-learning) — Microsoft Learn
EliExplains lessons are original prose written from the open, credible references above. See Copyright & Licensing.
Researched 2026-08-21
Educational content only. It is not medical, legal or professional advice. Found an error? Tell us.