Scikit-learn DBSCAN · Implementation Guide

from sklearn.cluster import DBSCAN
from sklearn.preprocessing import StandardScaler
from sklearn.datasets import make_blobs

# 1. Always standardise before DBSCAN (distance-based)
X, _ = make_blobs(n_samples=500, centers=4, n_features=2, random_state=42)
X = StandardScaler().fit_transform(X)

# 2. Fit
db = DBSCAN(eps=0.3, min_samples=5)
labels = db.fit_predict(X)

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

# 4. Access core sample indices
core_indices = db.core_sample_indices_

Key Attributes

• labels_: Cluster label for each point (-1 = noise)
• core_sample_indices_: Indices of core points (not border, not noise)
• components_: The actual core point instances (copy of X at core_sample_indices_)

Memory Considerations

DBSCAN requires an $O(n^2)$ or $O(n\text{d})$ distance computation depending on the algorithm. For n > 50,000, use algorithm='ball_tree' or algorithm='kd_tree' and leaf_size=30. For n > 500,000, consider approximate methods (HDBSCAN with approx_min_span_tree=True) or pre-computed distance matrices on a subset of features.

Common Errors

• All points labeled -1 (noise): ε is too small or minPts is too high. Increase ε or decrease minPts.
• Single large cluster + many noise points: ε is borderline — some points barely connected. Slightly increase ε or use HDBSCAN.
• Memory error on large dataset: Use ball_tree/kd_tree, reduce dimensionality with PCA first, or switch to HDBSCAN.