In plain words
Most AI learning starts with examples and right answers. Unsupervised learning skips the answers. The computer gets a heap of raw data, such as songs, shoppers or bank transactions, and nobody tells it what to look for.
Its job is to find structure on its own. Which items are similar? Which things tend to show up together? Which records look unusual? The most common result is a set of groups, called clusters.
The computer does not know what the groups mean. People look at them afterwards and give them names: weekend shoppers, bargain hunters, late night listeners. That makes unsupervised learning a tool for exploring data we do not understand yet.
Everyday example
Sorting a shoebox of old photos
Someone hands you 500 loose family photos with no dates and no notes. You start making piles: these look like one wedding, these are from the same beach, these share a school uniform. Nobody gave you the categories. You found them from what the photos have in common, and only at the end do you write a name on each pile.
Where the example breaks: You recognise a wedding because you know what weddings are. The computer only measures how alike items are in numbers, so its groups can be odd or meaningless, and a person still has to check them.
How it actually works
In unsupervised learning the training data has features but no labels, so there is no right answer to score against.
Turn items into numbers
Each item becomes a list of numbers: a song's tempo and loudness, a shopper's visits and average spend. Items that are alike end up close together in that number space.
Group what is close
Clustering methods such as k-means split items into groups whose members sit near each other. This is the most common unsupervised task.
Find what goes together
Other methods look for things that appear together. Google's ML glossary gives the example of shopping carts with lemons also often holding antacids, found by a method called principal component analysis.
Compress and spot oddities
Some methods squeeze many features into a few that keep most of the information. Others flag items that fit no group, which helps catch fraud and abuse.
Pass the result on
The groups often feed other systems, such as a music recommender, or are reviewed and named by people.
- Training data
- Features only, no labels
- Most common task
- Clustering: grouping similar examples
- Classic methods
- k-means clustering, principal component analysis (PCA), anomaly detection
- Used for
- Customer groups, music and shopping recommendations, fraud and abuse detection, exploring new data
Self-supervised learning, the way large language models begin, also uses unlabelled data, but it creates answers from the data itself and then learns as if supervised.
Questions people ask
What is the difference between supervised and unsupervised learning?
Supervised learning is shown the right answer for every example and learns to predict it. Unsupervised learning gets no answers and finds structure on its own. Supervised answers a question you already have; unsupervised helps you discover which questions are worth asking.
What is an example of unsupervised learning?
A music app grouping songs by how they sound, a shop sorting customers into types by how they buy, or a bank flagging transactions that look unlike everything else. In each case nobody labelled the data first.
Is clustering the same as unsupervised learning?
Clustering is the most common kind of unsupervised learning, but not the only one. Finding items that go together, compressing data into fewer features and spotting unusual records are unsupervised too.
Is ChatGPT trained with unsupervised learning?
Not exactly. Its base model learned from unlabelled text, but by predicting the next word, which turns the text itself into the answer key. That is called self-supervised learning, a close cousin of unsupervised learning.