Unsupervised

What it is

Most machine learning is taught with worked answers: here is a picture, here is the word for it. Unsupervised learning removes the answers. The program is handed only the measurements and asked what shape they make. Three questions come up again and again. Which items resemble each other closely enough to form a group, called a cluster? Where does the data crowd together and where is it sparse? And can many measurements per item be summarised by a few numbers without losing much, a step known as dimensionality reduction?

The machinery is mostly a notion of distance. Decide how to measure how unlike two items are, and the groupings follow. One common method, k-means, asks for a number of groups in advance, then shuffles items between groups until each sits with the ones nearest to it. Other methods build a network in which similar items are linked, and look for densely connected regions. Nothing in these procedures knows what a group means. The names and the interpretation are supplied afterwards, by a person.

Why scientists use it

Labels are expensive. Someone must decide what each item is, and in science that often means an expert, an instrument or a slow calculation. Many scientific datasets arrive unlabelled by their nature: a spectrometer or a telescope produces measurements, not verdicts. Unsupervised methods let researchers work with what they already have, at the scale they already have it.

The idea also stands in for a particular kind of human labour. A researcher faced with thousands of spectra, or a map of every mixture of three elements, would otherwise sit and sort them by eye, drawing boundaries where the patterns seem to change. Unsupervised methods do that sorting consistently and quickly. They are also used to tidy data before anything else happens: squeezing a continuous quantity into a handful of categories, or trimming many measurements down to a manageable few, so that later steps have less to chew on.

In the record

Limits

Finding groups is not the same as finding real ones. These methods will return a tidy answer from noise, and will happily split one thing in two or merge two into one, because nothing tells them what matters. Change the distance measure, the scaling of the measurements, or the requested number of groups, and the answer changes. The groups also arrive nameless: deciding that a cluster corresponds to a crystal phase, a kind of site or anything else is a judgement made outside the method, and it can be wrong.

Written by AIxSci: the background from general knowledge, the examples from checked records.

Every article