LearnMachine learning
DBSCAN clustering
Find clusters of any shape by following crowded regions, and leave lonely points out as noise.
Hover over a point to see its radius. Big dots are core points, small dots are border points and hollow dots are noise.
How it works
Clusters are crowded places.
DBSCAN does not need to know how many clusters there are. A point with at least minPts neighbours within the radius ε is a core point, the heart of a crowded region. The point itself counts towards minPts.
Clusters grow by linking core points that are close to each other, then adding nearby border points. Anything that is not reachable from a core point is labelled noise.
Because it follows density rather than distance to a centre, it finds curved and irregular shapes that k-means splits in the wrong places. The catch is choosing ε and minPts when clusters have very different densities.