What is Supervised Learning?
Supervised learning is training a model on examples that already carry the correct answer, so the model learns a mapping it can apply to new, unlabelled inputs.
The plain definition
In supervised learning you hand the algorithm pairs: an input and the correct output. Ten thousand emails, each tagged spam or not spam. Fifty thousand X-rays, each with a radiologist's finding. Three years of invoices, each with the account it was ultimately coded to.
The model's job is to find a function that reproduces those labels, and — this is the part that matters — keeps working on inputs it has never seen. The label is the supervision. It is also, almost always, the bottleneck.
Classification versus regression
Supervised problems split by what kind of thing you are predicting. Classification predicts a category: fraud or legitimate, one of twelve product tiers, which of forty ICD codes applies. Regression predicts a number: expected delivery days, next month's demand, a claim's likely cost.
The split is not cosmetic — it changes the loss function, the metrics, and how you interpret errors. A classifier is judged on precision, recall, and the shape of its confusion matrix. A regressor is judged on absolute or squared error. Getting these confused is a common way to ship a model that scores well and behaves badly.
The real cost is labelling
Teams routinely underestimate this. Algorithms are commodities; labelled data specific to your problem is not. If a domain expert takes ninety seconds per example and you need thirty thousand examples, that is roughly seven hundred and fifty hours of specialist time. On a clinical or legal problem, that expert is expensive and busy.
Several tactics reduce the bill. Transfer learning starts from a pretrained model so you need far fewer labels. Active learning asks the expert only for the examples the model is currently most uncertain about, which is typically where the information density is highest. Weak supervision derives noisy labels from heuristics and existing business rules, then models the noise.
One thing that does not work: labelling carelessly to hit a volume target. Label noise puts a hard ceiling on achievable accuracy, and a model trained on inconsistent labels will faithfully learn the inconsistency.
Leakage: the failure that looks like success
The most expensive mistake in supervised learning is target leakage — letting information into the features that would not exist at prediction time. A churn model that includes cancellation_reason will score beautifully in testing and be worthless in production, because in production that field is empty until after the customer has already churned.
Leakage also arrives through time. If you split records randomly on a time-ordered process, the model trains on the future and tests on the past. Any model that must run forward in time should be validated forward in time, with a strict temporal split.
- Labels are the supervision — each training example carries the correct answer
- Two shapes — classification predicts categories, regression predicts numbers
- Labelling dominates cost — expert time, not compute, is usually the binding constraint
- Guard against leakage — no feature may encode information unavailable at prediction time
More on Supervised Learning.
How much labelled data do I need for supervised learning?
It depends on how much signal each label carries and whether you are starting from scratch. Fine-tuning a pretrained language or vision model can produce useful results from a few hundred to a few thousand well-chosen examples. Training from scratch on a hard perceptual task can need six or seven figures. The honest answer is to label a small pilot set, plot accuracy against training-set size, and extrapolate the curve before committing budget.
What is the difference between supervised and unsupervised learning?
Supervised learning has labelled target values and learns to predict them. Unsupervised learning has no labels and instead finds structure — clusters, low-dimensional representations, density estimates. If you can state the exact answer you want for each training row, your problem is supervised.
Can I use supervised learning with imbalanced classes?
Yes, but accuracy stops being a meaningful metric. If one in a thousand transactions is fraudulent, a model that predicts 'never fraud' is 99.9% accurate and completely useless. Use precision, recall, and area under the precision-recall curve; consider class weighting or resampling; and pick the decision threshold from the business cost of a false positive versus a false negative rather than leaving it at 0.5.
Related terms.
Unsupervised Learning
Unsupervised learning finds structure in data that has no labels, producing groupings, compressed representations, or density esti…
ReadTraining Data
Training data is the set of examples a model learns from, and its coverage, label quality, and representativeness set a hard ceili…
ReadConfusion Matrix
A confusion matrix is a table of a classifier's predictions against actual outcomes, splitting results into true positives, false …
ReadOverfitting
Overfitting is when a model fits the particular noise and quirks of its training data so closely that its performance on new, unse…
ReadNeed this built?
Applying this to a
real system?
We build and ship production machine learning. If supervised learning is part of a problem you are working on, tell us about it.
Start the conversationor write to us at [email protected]