A spam filter flags an email as junk because it has seen thousands of messages already marked "spam" or "not spam." That is supervised learning in miniature: a model learns a mapping from inputs to outputs using examples where the correct answer is known. Each training pair carries a label, and the algorithm adjusts its internal parameters until its predictions match those labels closely enough to generalize to new data.

The approach splits into two broad families. Classification predicts a category, such as whether a tumor is benign or malignant. Regression predicts a number, such as tomorrow's temperature or a house price. Both rely on the same core loop: make a prediction, measure the error against the label, and nudge the weights to reduce that error.

Common algorithms

• Linear and logistic regression for simple, interpretable baselines
• Decision trees and random forests for tabular data
• Support vector machines for margin-based classification
• Neural networks for images, speech, and language

Labels cost money and time. A radiologist who annotates 10,000 scans is expensive; a crowd of workers labeling sentiment is slow and noisy. That expense is the main reason unsupervised and semi-supervised methods attract so much attention. When clean labels exist in abundance, though, supervised learning remains the most reliable tool in the box.