Supervised vs unsupervised learning

Last lesson you saw that machine learning means learning from examples. But examples come in two very different flavours: some come with the right answer attached, and some don’t. That single difference splits machine learning into its two biggest families.

Supervised learning: learning from right answers

In supervised learning, every example comes with the answer we want the computer to learn. We call the input $x$ and the answer, or label, $y$:

Input $x$	Output $y$	Application
an email	spam or not spam	spam filter
a photo of a skin spot	harmless or worth checking	medical screening
a sentence in English	the same sentence in French	translation
a house’s size	its sale price	price estimate

The model sees many $(x, y)$ pairs, learns the mapping from $x$ to $y$, and can then predict $y$ for an $x$ it has never seen. It’s called supervised because the labels act like a teacher checking every answer.

Supervised problems come in two main kinds:

• Regression predicts a number from a continuous range: a price, a temperature, minutes until your bus arrives. There are infinitely many possible answers.
• Classification predicts a category from a small set: spam or not spam, cat or dog or rabbit, which digit was handwritten. The answers are a short list.

The rest of this section is about regression. Classification comes in Section 3.

Unsupervised learning: finding structure without answers

Now imagine you run a shop and have data on hundreds of customers (how often they visit, how much they spend) but nobody has labelled them. There is no “right answer” column. You can still ask a useful question: are there natural groups in here?

That is unsupervised learning: the data has only inputs $x$, no labels $y$, and the algorithm’s job is to find structure on its own. The most common kind is clustering, grouping similar points together. Others include anomaly detection (spotting the odd one out, like an unusual card payment) and dimensionality reduction (squeezing big data into fewer numbers while keeping what matters).

See the difference

Below is the customer data. Toggle Show labels to switch between a supervised view (each customer’s group is known, shown by colour) and an unsupervised one (just grey dots). Then let the k-means algorithm find the groups by itself: it drops $k$ markers, called centroids, assigns every point to its nearest centroid, moves each centroid to the middle of its points, and repeats until nothing changes.

Loading the interactive graphic…

Play walkthroughStep 1 of 9

Each dot is a customer: how often they visit, and how much they spend. The colours are labels someone gave us, so this is the supervised view.

Things to try:

• Set $k$ to 2 or 5. The algorithm always finds exactly $k$ groups, even when that’s not the “true” number, so choosing $k$ is part of the job.
• Press New random start a few times with $k = 3$. Usually it finds the three groups, but a bad start can get stuck in a worse answer (a local minimum). Real systems run k-means several times and keep the best result.

Check yourself

Self-check · 4 questions

Supervised vs unsupervised learning

1. What makes a learning problem supervised?

It uses a neural networkIt has more than 1,000 examplesIt is watched by a human while it trainsEach training example comes with the correct answer y

2. Predicting tomorrow's maximum temperature in degrees is an example of…

RegressionClusteringAnomaly detectionClassification

3. A music service groups millions of songs into styles without anyone labelling them. Which kind of learning is this?

None, because this needs hand-written rulesSupervised regressionSupervised classificationUnsupervised clustering

4. In the k-means widget, what happens in each round after points are assigned to their nearest centroid?

Each centroid moves to the average position of its assigned pointsk increases by oneThe points move toward the centroidsThe labels are revealed

Check answers