Unsupervised learning finds structure in data that has no labels — no "correct answer" is provided, so the algorithm groups, compresses or otherwise organizes the data based on patterns in the features alone.
Two Main Tasks
| Task | Goal | Example Algorithms |
|---|---|---|
| Clustering | Group similar data points together | K-Means, Hierarchical Clustering, DBSCAN |
| Dimensionality Reduction | Compress many features into fewer, while keeping most of the information | PCA, t-SNE, UMAP |
Minimal Example — Customer Segmentation
from sklearn.cluster import KMeans
import numpy as np
# annual_spend (in thousands), visits_per_month — no labels given
X = np.array([[2, 1], [3, 1], [40, 8], [42, 9], [41, 7]])
model = KMeans(n_clusters=2, random_state=42, n_init=10)
labels = model.fit_predict(X)
print(labels)Expected output: something like [0 0 1 1 1] — the algorithm separated low-spend/low-visit customers from high-spend/high-visit customers purely from the numbers, with no "customer type" label ever provided.
Why It's Harder to Evaluate
Without ground-truth labels, there's no single "accuracy" to compute. Evaluation relies on internal metrics like silhouette score, or on whether the resulting groups make business sense to a human reviewer.
Practical Use Cases
- Customer segmentation for targeted marketing
- Anomaly/fraud detection (points that don't fit any cluster)
- Compressing high-dimensional data for visualization or as input to another model
- Topic discovery in a collection of documents
Advantages
- Doesn't require expensive labeled data
- Can reveal structure a human analyst wouldn't have thought to look for
Limitations
- No objective "correctness" measure — results need domain judgment to interpret
- Sensitive to feature scaling and the choice of number of clusters (see elbow method)
Common Mistakes
- Running clustering on unscaled features — a feature ranging 0–100000 will dominate distance calculations over one ranging 0–1; always apply feature scaling first.
- Assuming clusters found by an algorithm automatically correspond to meaningful real-world categories — they need to be validated against business context.
Interview Relevance
Q: "How would you evaluate a clustering result with no ground truth?" Use internal metrics (silhouette score, within-cluster sum of squares) alongside a domain-expert sanity check of what the clusters actually represent.
Practice Question
You're given transaction data with no fraud labels. Describe how you'd use unsupervised learning to flag potentially fraudulent transactions.
