DBSCAN clustering — density, not a cluster count
DBSCAN never asks for k. A row is a core point when at least minSamples neighbors sit inside radius eps. Clusters are chains of core points, border points hang off those chains, and everything else is labeled noise (−1). The shape can be a ring, a crescent, or a streak, because nothing about the update assumes a sphere.
On scaled Iris with minSamples 5 and eps estimated from the k-distance quantile, DBSCAN kept 2 clusters, silhouette 0.598, and adjusted Rand 0.552 against species. It separated the compact setosa cloud and treated the other two species as one denser band. On the wine chemistry table the same defaults kept a single dense region and marked 8 rows as noise, so silhouette is not defined there: one cluster has nothing to separate from.
eps is in the scaled units, not the original ones. If you skip scaling, a column with a wide range sets the radius and the rest of the table looks like a single point.
Settings that change this fit
- eps
- Neighborhood radius. Shrink it and chains break into noise. Grow it and separate clouds fuse.
- minSamples
- How many neighbors a core point needs. Raising it demands a denser region and usually increases the noise count.