DBSCAN

Finds clusters of any shape by following dense neighborhoods — and knows when a point is just noise.

Key Points

• Groups points based on density — areas with many nearby points form clusters, sparse areas don't.
• Automatically identifies outliers as 'noise' rather than forcing every point into a cluster.
• Doesn't require specifying the number of clusters in advance, only two parameters: eps (neighborhood radius) and min_samples.
• Can find non-spherical, irregularly shaped clusters that K-Means would fail on.
• Struggles when clusters have very different densities — one eps value can't fit all of them.

What it is

DBSCAN (Density-Based Spatial Clustering of Applications with Noise) groups points that are closely packed together into clusters and marks points that sit alone in low-density regions as noise/outliers. Unlike K-Means, it doesn't need a predefined number of clusters and can find clusters of any shape.

Why it exists

K-Means assumes clusters are round and roughly equal in size, and forces every point into some cluster even if it's really an outlier. DBSCAN exists to handle real-world data where clusters can be irregularly shaped and where some points genuinely don't belong to any group — like fraud transactions or sensor glitches.

Intuition

How it works

1. Pick eps and min_samples

   eps defines the radius of a neighborhood; min_samples defines how many points must be within that radius to count as 'dense.'

2. Classify each point

   A point with at least min_samples neighbors within eps is a 'core point.' Points near a core point but without enough neighbors themselves are 'border points.' Everything else is 'noise.'

3. Expand clusters

   Starting from a core point, grow the cluster by including all reachable core and border points connected through dense regions.

4. Label remaining points as noise

   Any point not reachable from a core point is marked as an outlier, not forced into a cluster.

When to use

• Use when clusters have irregular, non-spherical shapes.
• Use when your data legitimately contains outliers that shouldn't be forced into a cluster — e.g. anomaly detection.
• Use when you don't know the number of clusters in advance.

When not to use

• Do not use when clusters have very different densities — a single eps value won't work well for all of them.
• Do not use on very high-dimensional data without dimensionality reduction first — distance becomes less meaningful (curse of dimensionality).
• Do not expect it to work well without carefully tuning eps and min_samples — results are very sensitive to these.

Common mistakes

• Picking eps arbitrarily instead of using a k-distance plot to inform the choice.
• Applying DBSCAN to unscaled features, distorting what 'nearby' even means.
• Assuming every point must belong to a cluster — DBSCAN's noise label is a deliberate and often useful feature, not a bug.

In practice

Used for anomaly/fraud detection, geographic/spatial clustering (e.g. grouping GPS coordinates into hotspots), and any scenario where irregular cluster shapes or noise are expected in the data.

Related concepts