AIM-01-001-P · Primer — Supervised vs Unsupervised

The idea in one sentence

Supervised learning predicts a known answer you were handed, while unsupervised learning looks for structure in data that has no answer attached.

Why it exists

Almost every machine-learning problem starts at the same fork: do you already know the right answer for your examples, or are you exploring data to see what patterns fall out? The two branches use different algorithms, different data, and different ways of measuring success — so naming which one you're on is the first real decision, before you touch a single model. The folk explanation is "one has a teacher and one is on its own," which is close but misleading: the true dividing line is not how much guidance the algorithm gets, it's what question you are asking of the data.

How it works

Supervised learning starts with labeled examples — every row of data is paired with the correct answer, called the label. Give a model five thousand emails each already marked spam or not spam, and it learns the mapping from the words to that label; then on a brand-new email it has never seen, it predicts spam or not spam. Unsupervised learning has no labels at all: hand it ten thousand customers and their purchase histories with no right answer, and its job is to find structure on its own — a clustering algorithm might surface four natural groups of shoppers you never defined. Each branch has two common flavors: supervised splits into classification (predict a category, like spam) and regression (predict a number, like a house price at 420,000 dollars); unsupervised splits into clustering (group similar items) and dimensionality reduction (compress many features into a few).

See it — one fork: do your rows carry a label?

flowchart TD
    D["Your data rows"] --> Q{"Is each row<br/>paired with a<br/>known answer?"}
    Q -->|"Yes, labeled"| S["Supervised:<br/>predict that target"]
    Q -->|"No labels"| U["Unsupervised:<br/>find hidden structure"]
    S --> C["Classification<br/>predict a category"]
    S --> R["Regression<br/>predict a number"]
    U --> CL["Clustering<br/>group similar rows"]
    U --> DR["Dim. reduction<br/>compress features"]

The one gotcha

"No labels" does not mean "no supervision" in some vague, hand-wavy sense — it means a fundamentally different question. Supervised asks "predict this specific known target"; unsupervised asks "what structure is hidden in here at all." One is a prediction problem with a target column, the other is a discovery problem with no target — which is exactly why you can grade a supervised model against a right answer (accuracy, loss) but have no answer key to score an unsupervised one the same way. The fast test: do I have a target column I want to predict? If yes, supervised; if I'm just exploring, unsupervised.

See it (visualization plan)

Two side-by-side Cards/Tiles panels. Supervised: a Pipeline of labeled data → model learns mapping → predict the label, then the spam worked example (5,000 rows · label spam? · learns words→label). Unsupervised: a Layers panel of 10,000 customers, no label (dashed/amber) resolving into four discovered clusters. The gotcha lands as a Chain: "No labels" → is NOT "no supervision" → it is a different QUESTION, then two contrasting question cards (predict a known target vs find hidden structure).

The one-line takeaway

Labels decide the question: predict a known target, or find hidden structure.

Where you'll meet it

In scikit-learn, supervised estimators expose fit then predict (LogisticRegression, RandomForestClassifier), while clustering and PCA expose fit then transform (KMeans, PCA) — the API itself encodes the split. PyTorch and TensorFlow live mostly on the supervised side; the unsupervised ideas here lead straight into k-means clustering (AIM-01-045), PCA and dimensionality reduction (AIM-01-046), and embeddings (AIM-01-049).

Named concepts: supervised learning, unsupervised learning