Clustering groups data points so that points in the same group are more similar to each other than to points in other groups — with no labels given, no "correct" grouping told to the algorithm in advance. It's the primary technique of unsupervised learning.
Why Group Data With No Labels At All?
Sometimes there's no target variable to predict — you just want to understand a dataset's natural structure. "What are our distinct customer segments?" has no ground-truth answer sitting in a spreadsheet somewhere; clustering discovers a reasonable answer directly from how customers' features naturally group together.
The Three Main Clustering Approaches
| Approach | Core Idea | Full Note |
|---|---|---|
| Centroid-based | Group points around learned center points | K-Means |
| Hierarchical | Build a tree of nested clusters, merged step by step | Hierarchical Clustering |
| Density-based | Group points in dense regions, flag sparse regions as noise | DBSCAN |
Minimal Working Example
from sklearn.cluster import KMeans
import numpy as np
X = np.array([[2,2],[2,4],[2,6],[8,2],[8,4],[8,6]]) # two obvious groups
model = KMeans(n_clusters=2, random_state=42, n_init=10)
labels = model.fit_predict(X)
print(labels) # e.g. [0 0 0 1 1 1] -- the algorithm found the two groups with no labels givenWhy Clustering Is Harder to Evaluate Than Supervised Learning
Without ground-truth labels, there's no direct "accuracy" to compute. Evaluation instead relies on internal metrics — how tight and well-separated the resulting clusters are — covered in Clustering Evaluation and Silhouette Score — combined with a human sanity check of whether the groups actually make business sense.
Practical Use Cases
- Customer segmentation for targeted marketing
- Anomaly/fraud detection — points that don't fit any dense cluster stand out
- Document/topic grouping with no predefined categories
- Image compression — grouping similar colors into a smaller palette
Common Mistakes
- Running any distance-based clustering algorithm on unscaled features — exactly the same scale-sensitivity issue as KNN.
- Assuming clusters found by an algorithm automatically correspond to meaningful real-world categories — they need business-context validation, not just a good internal metric score.
- Picking one clustering algorithm without considering that different algorithms make very different structural assumptions — see the comparison in each specific algorithm's note.
Interview Relevance
Q: "How would you validate a clustering result with no ground truth to compare against?" Use internal metrics like silhouette score or within-cluster variance, alongside a domain-expert review of whether the resulting clusters represent something genuinely meaningful and actionable.
Practice Question
A retailer wants to identify distinct shopping behavior groups among 50,000 customers with no predefined categories. Would this be a supervised or unsupervised problem, and which family of clustering algorithm would you start with?
