DBSCAN vs k-means vs HDBSCAN · A Practical Comparison

Three clustering algorithms dominate practical work. Choosing between them requires understanding their assumptions.

The Assumptions

• k-means: Spherical, equal size · Forces into cluster · Yes
• DBSCAN: Arbitrary · Explicit noise labels · No
• HDBSCAN: Arbitrary · Explicit noise labels · No

When Each Wins

k-means wins when data is well-separated, spherical, and roughly equal cluster sizes. Fast: O(n) per iteration.

DBSCAN wins with arbitrary-shaped clusters, spatial data, and a known distance threshold (e.g., “within 100 metres”). Fails on varying density.

HDBSCAN wins as the default choice for real-world data. Single parameter, handles varying density, builds cluster hierarchy.

When All Fail

• High-dimensional data: distance concentration makes density meaningless. Reduce dimensions first.
• Radically different cluster sizes
• Overlapping clusters: use soft clustering (Gaussian Mixture Models)

Guideline

Start with HDBSCAN for exploratory clustering. Use k-means for speed and simplicity on well-behaved data. Use DBSCAN when domain knowledge provides a natural ε threshold.