Clustering
Grouping things that resemble each other — and the fact that the algorithm always returns groups, whether or not any exist.
When not to use it
- When you already know the categories. That's classification, and it's more accurate and measurable.
- On high-dimensional data without reduction. Distances concentrate and "similar" stops meaning anything.
- To justify a decision on its own. A cluster is a hypothesis. Someone has to look at it and vouch for it.
Reach for something else instead
- Classification when the groups are known and you have labels.
- Manual segmentation on business rules — often the right answer, and it has the advantage of being explicable.
- Dimensionality reduction plus looking — sometimes you just want to see the shape of the data, not commit to groups.
Sources & further reading
- Arthur & Vassilvitskii (2007), k-means++: The Advantages of Careful Seeding — why initialisation matters, and the fix.
- Ester et al. (1996), A Density-Based Algorithm for Discovering Clusters — DBSCAN, and clusters that aren't blobs.
- von Luxburg, Williamson & Guyon (2012), Clustering: Science or Art? — the stability and evaluation problem.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Trusting k from the elbow method. It's a hint. The right k is the one a domain expert can name.
- Running k-means once. It converges to local optima; different seeds give different answers, and that variation is information.
- Clustering unscaled data, then discovering the groups are entirely about revenue because revenue had the biggest numbers.
At a glance
FieldMachine Learning
Core ideagroup by similarity
Defaultk-means, with spherical assumptions
Better for odd shapesDBSCAN
Always returnsgroups, real or not
DifficultyBeginner → Intermediate
Flashcards for this concept
Question
—
Answer
—
1 / 4
Often compared with
k-means vs. DBSCAN — forcing every point into a round group vs. finding dense regions and calling the rest noise.