What is it?
You show a model many examples that already have the right answer — emails marked spam or not, houses with their sale prices — and it learns a rule that predicts the answer for new cases. "Supervised" refers to the answers acting as a teacher.
Also known as: supervised machine learning, classification and regression, labelled data learning, learning from labels, predictive modelling.
Technical definition
Given a dataset of pairs (x, y) drawn from an unknown distribution, supervised learning fits a function f_θ by minimising an empirical loss (cross-entropy for classification, squared error for regression, etc.) plus optional regularisation, aiming to minimise expected loss on unseen data. Evaluation uses held-out validation and test splits; cross-validation helps with small datasets. Failure modes include label noise, distribution shift and leakage between train and test.
Entity type: concept · Freshness class: stable.
When should I use it?
Use when you have (or can obtain) reliable labelled examples and a clear prediction target.
Implementation patterns
- Split data into train / validation / test before any preprocessing that learns from data.
- Start with a simple baseline (linear model, tree) before a neural network.
- Track metrics suited to the problem (precision/recall for imbalance).
- Use cross-validation for small datasets.
Technologies and frameworks
Commonly associated: scikit-learn, PyTorch, XGBoost / gradient boosting, pandas. Names are examples of what exists, not recommendations, and projects change quickly.
Practical guide
- Define the target and the metric.
- Collect and clean labelled data; audit labels.
- Split before preprocessing.
- Train a baseline, then iterate.
- Evaluate once on the held-out test set; monitor after launch (Model Drift and Production Monitoring).
Common failure modes
- Data leakage: information from the test set influences training.
- Label noise or inconsistent labelling.
- Class imbalance hidden by accuracy.
- Training data not representative of production inputs.
Troubleshooting
- Great test score, poor production results — Check leakage and distribution shift between test and production data.
- Accuracy high but rare class missed — Use precision/recall/F1, resampling or class weights.
- Unstable results — Use cross-validation and fix random seeds for comparison.
Contrasted with
- Unsupervised Learning — No labels; finds structure.
- Reinforcement Learning — Learns from rewards after actions, not labelled answers.
- Self-Supervised Learning — Creates its own labels from the data.
Questions this page answers
- What is supervised learning?
- What is the difference between classification and regression?
- What is the difference between supervised and unsupervised learning?
- How do I split data for training and testing?
Typical goals: understand the basic ML setup; choose between supervised and unsupervised learning; avoid train/test leakage.
Sources and verification
Status: not independently verified. This entry was authored on 2026-10-06 from general technical knowledge, without live checking of the sources below. Treat every source as a lead to confirm (identifier, title, authorship, current version) before publication or citation.
Freshness class: stable — Core ideas that change slowly (maths, classic algorithms). Re-review yearly.
Canonical sources (unverified leads):
- Hastie, Tibshirani & Friedman, The Elements of Statistical Learning (2nd ed.) — hastie.su.domains/ElemStatLearn
- Bishop, Pattern Recognition and Machine Learning (2006) — locator to be identified
- scikit-learn user guide — scikit-learn.org