What does Unsupervised Learning mean?

Unsupervised learning describes methods that find structure in data without a correct outcome attached to the examples. The method receives only the inputs and independently searches them for groups, relationships, or outliers. In supervised learning, the answer to every example is known in advance; here, nobody specifies what to look for in the first place.

Cluster analysis groups similar data records together. The k-means method needs the number of groups as an input, while DBSCAN derives it from the density of points and can also detect irregular shapes. Dimensionality reduction methods such as principal component analysis condense many features into a few, making the data easier to visualize. Outlier detection flags data records that differ noticeably from all the others.

Unsupervised learning pays off where a large dataset exists that nobody has classified yet. Typical cases include segmenting customers or applicants, spotting unusual transactions, and pre-sorting large document collections. If the desired classification is already known, however, a supervised method is the more direct and accurate path.

The advantage over manual classification is that no annotation by subject-matter experts is required. The methods work on the existing dataset and deliver an initial classification within hours. The groups found often then serve as the basis for a supervised model.

Before a discovered group is put to further use, someone with subject-matter knowledge needs to check what it is based on. When segmenting applicants, a method might split by postal code because that is where the largest differences lie, even though where someone lives is irrelevant to the task. Without this check, a coincidental similarity ends up baked into the downstream model.