What unsupervised learning is
Unsupervised learning is machine learning on data that has no labels. There is no correct answer attached to each example and nothing for the model to predict. Instead, the algorithm’s job is to discover hidden structure in the data on its own — which points are similar, how the data is distributed, or how it can be compressed. Because most of the world’s data is unlabelled, these techniques are essential for exploring datasets and preparing them for other tasks.
Clustering
The most common unsupervised task is clustering: grouping data points so that items in the same group are more alike than items in different groups.
- k-means assumes you want k groups, places k centre points, and iteratively assigns each item to its nearest centre and moves the centres to the middle of their members. It is fast but you must pick k in advance.
- DBSCAN groups points that are densely packed together and labels sparse outliers as noise. It can find clusters of arbitrary shape and does not need you to specify the number of clusters.
Clustering powers customer segmentation, document grouping, and anomaly detection.
Dimensionality reduction
Real datasets often have hundreds or thousands of features, which is hard to visualise or model. Dimensionality reduction compresses them into a handful of values while preserving the important structure.
- PCA (Principal Component Analysis) finds the directions along which the data varies most and projects onto them, keeping the largest sources of variation.
- t-SNE and similar methods produce 2-D maps where nearby points were nearby in the original space, making clusters visible to the human eye.
These methods help with visualisation, noise removal, and speeding up later training.
Generative models and autoencoders
Unsupervised learning also includes generative approaches that learn the underlying distribution of the data so they can produce new, similar examples.
- Autoencoders squeeze an input down to a compact code and then reconstruct it. By forcing the data through a narrow bottleneck, the network learns an efficient representation, useful for compression and anomaly detection.
- Broader generative families model the data distribution directly so they can sample fresh examples that resemble the training set.
Why it matters
Because there are no labels, unsupervised learning is harder to evaluate — there is no accuracy score to compute against a ground truth. Its value usually shows up downstream: cleaner features for a supervised model, an insightful segmentation, or a compressed representation. Many of the biggest modern models blur the line by generating their own labels, a hybrid approach covered under self-supervised learning.