Unsupervised Learning

Unsupervised learning finds structure in data that has no labels attached. Nobody tells the algorithm what the right answer is, because there is no right answer to give it.

You use it when you do not know what you are looking for. It can tell you that your customers fall into five distinct groups. It cannot tell you what those groups mean.

Think of it like this. Think of sorting a box of unfamiliar objects with no instructions. You would still end up with piles, grouped by whatever similarities you noticed. Naming the piles is a separate job.

How it works

• No labels go in. Only the raw data.
• Similarity does the work. The algorithm measures how alike examples are and groups accordingly.
• Structure comes out. Clusters, a lower-dimensional view, or a list of anomalies.
• A human interprets it. The algorithm produces groups. It has no idea what they represent.

Types

• Clustering. Group similar examples together, as in customer segmentation.
• Dimensionality reduction. Compress many features into a few while keeping the important variation.
• Anomaly detection. Find the examples that do not fit any pattern, which is how a lot of fraud detection starts.
• Association. Find things that co-occur, the classic being items bought together.

Common mistakes

• "The clusters are correct." They are one valid grouping. Change the algorithm or the number of clusters and you get a different, equally valid one.
• "The clusters are meaningful." They are mathematically similar. Whether that similarity matters to your business is a human judgement.
• "It should match my known categories." If you already have categories, you want supervised learning.