Clustering Skill

Discover hidden patterns and groupings in unlabeled data.

Quick Start

from sklearn.cluster import KMeans
from sklearn.preprocessing import StandardScaler
from sklearn.metrics import silhouette_score

# Always scale before clustering
scaler = StandardScaler()
X_scaled = scaler.fit_transform(X)

# Cluster
kmeans = KMeans(n_clusters=5, random_state=42, n_init=10)
labels = kmeans.fit_predict(X_scaled)

# Evaluate
score = silhouette_score(X_scaled, labels)
print(f"Silhouette Score: {score:.4f}")

Key Topics

1. Clustering Algorithms

Algorithm	Best For	Key Params
K-Means	Spherical clusters	n_clusters
DBSCAN	Arbitrary shapes, noise	eps, min_samples
Hierarchical	Nested clusters	linkage
HDBSCAN	Variable density	min_cluster_size

from sklearn.cluster import KMeans, DBSCAN, AgglomerativeClustering
import hdbscan

algorithms = {
    'kmeans': KMeans(n_clusters=5, random_state=42),
    'dbscan': DBSCAN(eps=0.5, min_samples=5),
    'hierarchical': AgglomerativeClustering(n_clusters=5),
    'hdbscan': hdbscan.HDBSCAN(min_cluster_size=15)
}

2. Finding Optimal K

from sklearn.metrics import silhouette_score, calinski_harabasz_score
import matplotlib.pyplot as plt

def find_optimal_k(X, max_k=15):
    """Find optimal number of clusters."""
    metrics = {'inertia': [], 'silhouette': [], 'calinski': []}
    K = range(2, max_k + 1)

    for k in K:
        kmeans = KMeans(n_clusters=k, random_state=42, n_init=10)
        labels = kmeans.fit_predict(X)

        metrics['inertia'].append(kmeans.inertia_)
        metrics['silhouette'].append(silhouette_score(X, labels))
        metrics['calinski'].append(calinski_harabasz_score(X, labels))

    # Find optimal k
    optimal_k = K[np.argmax(metrics['silhouette'])]
    return optimal_k, metrics

3. Dimensionality Reduction

Method	Preserves	Speed
PCA	Global variance	Fast
t-SNE	Local structure	Slow
UMAP	Both local/global	Fast

from sklearn.decomposition import PCA
from sklearn.manifold import TSNE
import umap

# PCA for preprocessing
pca = PCA(n_components=0.95)  # Keep 95% variance
X_pca = pca.fit_transform(X)

# UMAP for visualization
reducer = umap.UMAP(n_components=2, random_state=42)
X_2d = reducer.fit_transform(X)

4. Anomaly Detection

from sklearn.ensemble import IsolationForest
from sklearn.neighbors import LocalOutlierFactor

# Isolation Forest
iso_forest = IsolationForest(contamination=0.1, random_state=42)
anomalies = iso_forest.fit_predict(X)  # -1 for anomaly

# Local Outlier Factor
lof = LocalOutlierFactor(n_neighbors=20, contamination=0.1)
anomalies = lof.fit_predict(X)

5. Cluster Validation

from sklearn.metrics import (
    silhouette_score,
    calinski_harabasz_score,
    davies_bouldin_score
)

def validate_clustering(X, labels):
    """Comprehensive cluster validation."""
    return {
        'silhouette': silhouette_score(X, labels),  # Higher = better
        'calinski_harabasz': calinski_harabasz_score(X, labels),  # Higher = better
        'davies_bouldin': davies_bouldin_score(X, labels),  # Lower = better
        'n_clusters': len(set(labels) - {-1})
    }

Best Practices

DO

• Always scale features before clustering
• Try multiple algorithms
• Use multiple validation metrics
• Visualize clusters in 2D
• Interpret clusters with domain knowledge

DON'T

• Don't use K-Means for non-spherical clusters
• Don't ignore the silhouette score per cluster
• Don't assume first result is optimal
• Don't use t-SNE for new point projection

Exercises

Exercise 1: Elbow Method

# TODO: Implement elbow method to find optimal K
# Plot inertia vs K and identify elbow point

Exercise 2: Compare Algorithms

# TODO: Compare K-Means, DBSCAN, and HDBSCAN
# on the same dataset using silhouette score

Unit Test Template

import pytest
import numpy as np
from sklearn.datasets import make_blobs

def test_clustering_finds_groups():
    """Test clustering finds expected number of clusters."""
    X, y_true = make_blobs(n_samples=100, centers=3, random_state=42)

    kmeans = KMeans(n_clusters=3, random_state=42)
    labels = kmeans.fit_predict(X)

    assert len(set(labels)) == 3

def test_scaling_improves_score():
    """Test that scaling improves clustering quality."""
    X, _ = make_blobs(n_samples=100, centers=3)
    X[:, 0] *= 100  # Make first feature much larger

    # Without scaling
    labels_raw = KMeans(n_clusters=3).fit_predict(X)
    score_raw = silhouette_score(X, labels_raw)

    # With scaling
    X_scaled = StandardScaler().fit_transform(X)
    labels_scaled = KMeans(n_clusters=3).fit_predict(X_scaled)
    score_scaled = silhouette_score(X_scaled, labels_scaled)

    assert score_scaled > score_raw

Troubleshooting

Problem	Cause	Solution
All in one cluster	Wrong eps/K	Reduce eps or increase K
Too many clusters	Parameters too sensitive	Increase eps or min_samples
Poor silhouette	Wrong algorithm	Try different clustering method
Memory error	Large dataset	Use MiniBatchKMeans

Related Resources

• Agent: 03-unsupervised-learning
• Previous: supervised-learning
• Next: deep-learning

---

Version: 1.4.0 | Status: Production Ready