HDBSCAN clustering — stable clusters at several densities
HDBSCAN builds the same density connectivity as DBSCAN, then condenses it into a tree and keeps the branches that persist across a range of distance scales. A tight group and a loose group can both survive. Rows that never join a stable branch are noise. You set a minimum cluster size, not a radius and not k.
With min_cluster_size 5 on scaled Iris, the stable extraction kept 2 clusters, silhouette 0.590, adjusted Rand 0.564 against species. That is the same story DBSCAN told: setosa is stable, and the other two species are not stable as separate branches at that minimum size. Raising the minimum size toward 15 on a 150-row table will erase any branch smaller than that.
Do not read a noise label as an error. It means the point fell off every branch that lasted. If you need a label for every row, this is the wrong tool.
Settings that change this fit
- min_cluster_size
- Smallest branch that is allowed to count as a cluster. It is a size, not a distance, so it transfers more cleanly across columns than eps.
- min_samples
- How conservative the underlying density estimate is. Leave it equal to min_cluster_size unless the tree is sprouting tiny unstable leaves.