Supervised learning is the family of machine-learning algorithms that learns from labelled data: every training example comes paired with a correct answer. This image shows an owl, this patient has heart disease, this email is spam. The model learns a function that takes an input and predicts the correct answer for it.
Two main types of supervised task, distinguished by what kind of answer the model produces:
- Regression — the answer is a continuous numerical value. Given a person’s age and weight, predict their blood pressure. Given a year, predict the inflation rate.
- Classification (ML) — the answer is a discrete category. Is this email spam or not? Is this wine high quality or low? Does this patient have diabetes?
In both cases, the training procedure is roughly the same: pick a model with adjustable parameters, define a Loss function that measures how badly the model’s predictions agree with the labels, and use Gradient descent (or a closed-form solver when one exists) to find parameters that minimize the loss.
The other two families of machine learning are different in what feedback the algorithm gets:
- Unsupervised learning uses unlabelled data — just inputs, no correct answers. The model finds structure on its own: clusters, patterns, anomalies, low-dimensional representations. PCA and k-means clustering are unsupervised.
- Reinforcement learning is interactive — an agent takes actions in an environment, gets rewards or penalties, and learns a policy that maximizes cumulative reward. Chess engines, Go engines, and game-playing agents are trained this way.
A simple feel for the distinction: imagine someone shows us a stack of fruit photographs, some oranges and some apples. If the photographs are labelled (the orange ones say “orange”) we can learn to classify a new one. That’s supervised. If they’re unlabelled, we can still sort them into piles by visual similarity, color and shape, and end up with two piles even without knowing what they’re called. That’s unsupervised.
The Introduction to Data Science textbook focuses on supervised learning: linear and polynomial Regression for continuous outputs, Logistic regression for binary classification, trained with Gradient descent on standard losses (Mean squared error for regression, Binary cross-entropy for classification).