DBSCAN — Module 5.3

How This Lesson Fits the Module

K-Means assumes spherical, evenly sized groups. DBSCAN (Density-Based Spatial Clustering of Applications with Noise) finds clusters as dense regions separated by sparse space—and explicitly labels outliers as noise.

Use DBSCAN when shapes are irregular, cluster count is unknown, or you need a “no cluster” category for anomalies. It is the go-to density method before trying hierarchical approaches.

Learning Objectives

By the end of this lesson, students should be able to:

Define core points, border points, and noise in DBSCAN.
Tune eps and min_samples with domain and k-distance plots.
Implement DBSCAN in sklearn and interpret label −1 as noise.
Compare DBSCAN to K-Means on non-spherical data.
Explain why feature scaling strongly affects eps.
Choose DBSCAN when outlier detection and arbitrary shapes matter.

Core Concepts

DBSCAN grows clusters from core points—observations with at least min_samples neighbors within distance eps. Border points sit within eps of a core point but lack enough neighbors themselves. Everything else is noise (label −1).

Term	Definition
Core point	≥ `min_samples` points within radius `eps`
Border point	Within `eps` of a core point but not core itself
Noise	Not reachable from any core point; labeled −1
Density-reachable	Chain of core points within `eps` connects two observations

sklearn Implementation

DBSCAN does not have a predict method for new points in older sklearn versions—it is transductive (fits and labels the given set). Scale features so eps is interpretable in standardized space.

import numpy as np
from sklearn.cluster import DBSCAN
from sklearn.preprocessing import StandardScaler
from sklearn.neighbors import NearestNeighbors

X_scaled = StandardScaler().fit_transform(X)

# k-distance plot: distance to k-th neighbor (k = min_samples)
k = 5  # often set min_samples = 2 * dimensionality as a starting rule
nn = NearestNeighbors(n_neighbors=k)
nn.fit(X_scaled)
distances = np.sort(nn.kneighbors(X_scaled)[0][:, -1])

db = DBSCAN(eps=0.45, min_samples=5, metric="euclidean")
labels = db.fit_predict(X_scaled)

n_clusters = len(set(labels)) - (1 if -1 in labels else 0)
n_noise = (labels == -1).sum()
print(f"Clusters: {n_clusters}, Noise: {n_noise}")

df["cluster"] = labels
df[df["cluster"] == -1]  # flagged anomalies

Engineering Habit — k-Distance Plot for eps

Plot sorted distances to the k-th nearest neighbor. The “knee” where the curve bends upward suggests a reasonable eps. Combine with silhouette on non-noise points and business review of the noise fraction.

Tuning eps and min_samples

Parameter	Effect if too small	Effect if too large
`eps`	Many tiny clusters; most points become noise	Everything merges into one cluster
`min_samples`	More core points; fragile, fragmented clusters	Fewer cores; stricter density requirement

When DBSCAN Shines

Arbitrary cluster shapes (rings, blobs)
Unknown number of clusters
Built-in noise / anomaly labels
Geospatial or embedding neighborhoods

When to Avoid DBSCAN

Clusters differ greatly in density
Very high dimensions without reduction
You need fast scoring on streaming new points
Uniform global density (K-Means may suffice)

Critical Mistake — Copying eps Across Datasets

eps is not portable. After changing features, scaling, or sample size, re-tune from a k-distance plot. An eps that worked on 10k standardized rows may fail on 1M raw features.

DBSCAN vs K-Means vs Hierarchical

Method	Cluster count	Outliers	Shape assumption
K-Means	Fixed k	Forced into a cluster	Spherical, similar variance
DBSCAN	Discovered	Label −1	Density-connected regions
Hierarchical	Cut dendrogram	Manual or distance threshold	Nested structure visible

Knowledge Check

Short Answer: What label does DBSCAN assign to noise? Answer: −1.
True/False: DBSCAN requires you to specify the number of clusters. Answer: False—you specify density parameters, not k.
Multiple Choice: Best tool to pick eps: (a) elbow on inertia, (b) k-distance plot, (c) accuracy score. Answer: (b).
Short Answer: What makes a point a core point? Answer: At least min_samples neighbors within distance eps.
Short Answer: Why scale before DBSCAN? Answer: eps is a fixed radius; unequal feature scales distort distance.
True/False: DBSCAN can find non-spherical clusters. Answer: True.
Multiple Choice: Border points: (a) are noise, (b) are density-reachable from a core point but not themselves core, (c) define k. Answer: (b).
Short Answer: What happens if eps is too small? Answer: Most points become noise or tiny fragments.
True/False: sklearn DBSCAN has a simple predict for new points like K-Means. Answer: False—standard DBSCAN does not assign new points without extra logic.
Multiple Choice: High-dimensional data makes eps: (a) easier, (b) harder due to distance concentration, (c) irrelevant. Answer: (b).

Key Takeaways

DBSCAN finds density-connected clusters and flags sparse points as noise.
Tune eps with k-distance plots; set min_samples with dimensionality in mind.
Scale features; re-tune when the feature space changes.
Prefer K-Means for fixed, spherical segments; DBSCAN for irregular shapes and anomalies.
Next: Hierarchical Clustering—build and cut cluster trees.

Trainer’s Guide

Hands-on idea: Use sklearn’s make_moons dataset. Show K-Means failure vs DBSCAN success, then sweep eps to demonstrate over-merging and over-fragmentation.

Discussion prompt: Fraud analysts want noise points reviewed daily. What min_samples trade-off balances false alarms vs missed fraud?

Recap: DBSCAN finds density-connected clusters and labels sparse points as noise without choosing k. Continue with Hierarchical Clustering.