Bisecting K-Means — split the largest cluster
Bisecting K-Means starts with every row in one cluster and repeatedly runs a 2-means split. The cluster chosen for the next split is the one with the largest inertia, so a big loose group is broken before a small tight one. After k−1 successful splits you have k clusters. The path is a tree, but each cut is a K-Means cut, not a linkage merge.
On scaled Iris at k=3 the split silhouette was 0.457 and adjusted Rand against species was 0.641, within a hair of ordinary K-Means (0.460 and 0.620). The interesting difference shows up when one group is much larger than the others: ordinary K-Means can park two centroids inside the large group, while the biggest-inertia rule is pushed to divide that group first.
The result still has no noise label and still assumes each split is into two roughly spherical pieces. A ring-shaped cloud will be sliced through the middle.
Settings that change this fit
- n_clusters
- How many times the table is split, plus one. Each added cluster comes from bisecting an existing one.
- bisecting strategy
- Biggest inertia picks the loosest cluster. A biggest-size rule would instead always divide the most populous one, even if it is already tight.