MMOOZZGG.com

Glossary

What is clustering in machine learning

In short

Clustering in machine learning is a method that splits objects into groups of similar ones by itself, with no answers set in advance: customers, products, the texts of requests.

Machine learning · clustering

MMOOZZGG.com editorsChecked: 1 min read

For a child

Pour a big box of buttons onto the carpet and try to sort them, even though nobody told you what piles there should be. You decide yourself: here are the small white ones, there the big ones with four holes, and here the shiny ones. A computer can also sort things into piles without hints. A person then makes up names for the piles.

In plain words

Here the computer has no right answers. An algorithm gets, for example, a list of buyers with their purchases, order frequency and average check, and looks for who is similar to whom. People with close behavior land in one group, distant ones in different groups. The result is groups such as “rare large orders”, “regular discount lovers” and “lost customers”. A person decides what each group means and what to do with it. Similar complaints, news or product photos are grouped the same way.

For business

The method is needed when there is a lot of data and no ready split: customer segments for mailings, new topics in support requests, products with similar demand, unusual transactions. The result is checked with money: does a separate offer to each segment bring more sales than a common one? The cost is low if the data is already in one system; most of the spending goes to export, cleaning and an analyst’s reading of the groups. The risk is that the algorithm finds groups even where there are none in essence, and a nice picture leads the team the wrong way. Before launch, agree on what decision you will make for each group, or the report will stay just a report. The segments are then tested on one sales channel.

For a specialist

Clustering is an unsupervised learning task: split objects into groups so that within a group they are closer to each other than to objects of other groups. The k-means method needs the number of clusters set in advance and minimizes the sum of squared distances to the centroids; it tends to get stuck in local minima, so it is run from different starting points. Agglomerative hierarchical clustering builds a tree of merges. The density methods DBSCAN (Ester et al., 1996) and HDBSCAN treat an area of high point density as a cluster and mark single points as noise. Gaussian mixtures give soft assignment with probabilities. Without reference labels, quality is measured with internal metrics such as the silhouette coefficient; with labels, the adjusted Rand index is used. Texts are turned into embeddings before grouping.

Example

An online school split its students by behavior and found a group that buys a course and drops it in the second week. For them, it launched short reminders and homework reviews, and kept the old emails for everyone else.

Related terms

Questions and answers

How do you choose the number of clusters? In k-means, the number of groups is set in advance, so you try several options and look at the silhouette score. The business has the last word: the groups must be clear and call for different actions.
Where is clustering used in business? In customer segmentation, in finding new topics among requests, in grouping products and in finding suspicious transactions. Anywhere there is a lot of data and no ready categories.

Sources

Not sure whether your company needs this?

AI solutions map
Contact us