階層的クラスタリングとDBSCAN

k-Means assumes round, similar-sized clusters and a fixed k. When that breaks, other algorithms handle trickier shapes and densities. 🔍

Hierarchical clustering merges the two closest groups over and over, building a tree of nested clusters from the bottom up.

That tree is drawn as a dendrogram. Cut it at a chosen height and the branches below become your clusters.

A real perk is that you do not commit to k early. You inspect the dendrogram, then decide where to cut it.

How distance between groups is measured depends on the linkage, such as ward, average, or complete.

DBSCAN takes a different angle: it treats clusters as dense regions of points separated by sparser gaps.

DBSCAN needs a neighborhood radius eps and a minimum point count. Together they define what dense enough means.

Unlike k-Means, DBSCAN figures out the number of clusters itself from the density of your data.

Points in sparse areas are flagged as noise with the label minus one, instead of being forced into a cluster.

Because it follows density, DBSCAN can trace long, curved, or oddly shaped clusters that k-Means would slice apart.

Use hierarchical for nested structure you want to explore, and DBSCAN for irregular shapes with built-in outlier handling.

Let us confirm a standout trait of DBSCAN.

Hierarchical clustering builds a dendrogram you cut, while DBSCAN finds dense, any-shaped clusters and flags noise. 🎯