Module 2 · Lesson 4
What if we have a large dataset but no answer key?
That is where unsupervised learning becomes useful. Instead of learning to predict a supplied target for each example, the algorithm looks for structure in the data itself.
Three common jobs
Clustering groups data points that are similar according to a chosen representation and distance measure. A retailer might discover groups of customers with different purchasing patterns without first defining those groups by hand.
Dimensionality reduction compresses a dataset into fewer dimensions while trying to preserve important structure. Techniques such as PCA can make complex data easier to visualize or model.
Anomaly detection looks for observations that differ substantially from common patterns. That can be useful in areas such as equipment monitoring, fraud investigation, and quality control.
Raw observations
↓
measure relationships
↓
clusters / lower-dimensional structure / unusual points“No labels” does not mean “no human judgment”
It is tempting to say unsupervised learning requires zero human effort. That is too strong. People still decide what data to collect, how to represent it, what algorithm to use, how many clusters may be useful, what constitutes an anomaly, and whether the discovered structure has real-world meaning.
The important distinction is that the algorithm is not given a target label for every training example.
Where self-supervised learning fits
Self-supervised learning is related, but it deserves its own category. Instead of searching for structure without a prediction target, it creates a target from the data itself. That distinction becomes central when we reach language-model pre-training.
The next lesson focuses on self-supervised learning and shows exactly how raw text can generate its own training signal.
The human still has to interpret the result
A cluster is not automatically a business category. An outlier is not automatically fraud. A compressed dimension is not automatically meaningful to a person.
Unsupervised methods can reveal structure we did not know to look for, but the discovered patterns still need interpretation, validation, and context.