DBSCAN — Density-Based Clustering — Knowledge Check

Question 1

DBSCAN defines a cluster as:

Accepted Answer

A continuous region of high density — Which is why it handles arbitrary shapes, where k-means insists on roughly spherical clusters around centroids.

Question 2

An instance is a core instance when:

Accepted Answer

It has at least min_samples instances within ε, including itself — Core instances are the ones located in dense regions. Note that the count includes the instance itself.

Question 3

The region within distance ε of an instance is called its:

Accepted Answer

ε-neighborhood — Two parameters define the whole algorithm: the radius ε and the count min_samples.

Question 4

How do long, curved clusters end up as one cluster rather than many?

Accepted Answer

A chain of neighbouring core instances forms a single cluster — All instances in a core instance's neighbourhood join its cluster, and that neighbourhood can contain other core instances, so membership propagates transitively.

Question 5

An instance that is not a core instance and has no core instance in its neighbourhood is:

Accepted Answer

Considered an anomaly — DBSCAN is one of the few clustering algorithms with an explicit notion of a point belonging to nothing.

Question 6

In scikit-learn, anomalies appear in labels_ with which value?

Accepted Answer

−1 — They stay in the array, so labels_ always lines up with the input rows.

Question 7

On the moons dataset, DBSCAN with eps=0.05 and min_samples=5 produced:

Accepted Answer

Seven clusters plus a lot of anomalies — Géron's reaction is "How disappointing!" — the neighbourhoods are too small to reach along the crescents.

Question 8

Raising eps from 0.05 to 0.2 on that dataset gave:

Accepted Answer

A clustering that looks perfect — One parameter, same algorithm, same data — the difference between useless and correct.

Question 9

DBSCAN notably lacks which method?

Accepted Answer

predict() — It has fit_predict() but genuinely cannot assign a new instance to a cluster.

Question 10

Why did the authors leave that method out?

Accepted Answer

Because different classification algorithms suit different tasks, so the user should choose — And it is not hard to implement — the book trains a KNeighborsClassifier on the core instances.

DBSCAN — Density-Based Clustering

DBSCAN on two moons

Core, border, or noise?

Test your knowledge

DBSCAN defines a cluster as:

Learn these first

Interview Questions

References & Further Reading

Love this visualization?