MMOOZZGG.com

Glossary

What is supervised learning?

In short

Supervised learning is a way to train a model on examples in which every input already has the right answer attached.

AI basics · supervised learning

MMOOZZGG.com editorsChecked: 2 min read

For a child

Imagine a stack of cards with animals. On the back of each one is a name: cat, fox, raccoon. You look at the picture, guess, turn the card over and check yourself. After a hundred cards, you recognize a raccoon even in a new picture. A computer learns the same way: it gets many examples, and each one already has the right answer attached. The person who wrote the names on the cards is the teacher.

In plain words

You give the model a table of examples with both the question and the answer. Say, thousands of emails, and for each one you know whether it is spam. The model guesses the answer, checks it against the label and adjusts its settings to make fewer mistakes. This repeats many times. Then you bring it a new email without a label, and it decides by itself which folder to put it in. If the answer is a category (spam or not), the task is called classification. If the answer is a number, such as an apartment price or a delivery time, the task is called regression.

For business

Most applied AI in companies works on this scheme: customer churn forecasts, application scoring, the chance a deal closes, routing requests to departments. The main cost here is not the model but labeling: someone has to add the right answers to thousands of examples. Often they already sit in the CRM and the accounting system, such as the deal status or a product return. The risks: the model repeats the old mistakes and biases of employees, and history goes out of date over time. Before launch, check how much labeled history you have and how far you can trust it. A process review in a map of AI solutions shows where such a model pays off.

For a specialist

Supervised learning builds a function that predicts a target variable y from features x, using a training set of known pairs. The algorithm minimizes a loss function: cross-entropy for classification, mean squared error for regression. Quality is measured on a held-out test set, because the goal is not to memorize examples but to generalize the pattern to new ones. The scikit-learn user guide lists the main families: linear models, support vector machines (SVM), decision trees, ensembles (random forest, gradient boosting), nearest neighbors, naive Bayes and the multilayer perceptron. It also describes semi-supervised learning, where only part of the data is labeled. The theory is in the textbook by Hastie, Tibshirani and Friedman, “The Elements of Statistical Learning”, second edition, 2009. Typical problems: overfitting, target leakage into the features and distribution shift after launch.

Example

An online store takes two years of order history, where for each order it knows whether the buyer returned the product. The model learns from this history and flags new orders with a high return risk, so a manager can call the customer before shipping.

Related terms

Questions and answers

How is supervised learning different from unsupervised learning? In the first case, each example has a right answer, and the model learns to predict it. In the second, there are no answers, and the model looks for groups and patterns in the data on its own.
Where do you get labeled data for a model? Often it is already in your systems: the deal status in the CRM, the payment record in the accounting system (1C in Russia), the request category in support. If there are no labels, employees or labeling contractors add them, and this is the main cost of the project.

Sources

Not sure whether your company needs this?

AI solutions map
Contact us