Unsupervised Learning - Finding Hidden Patterns

So far in this series, we’ve worked with labeled data - every training example had a known answer. But what happens when you have data with no labels at all?

Welcome to unsupervised learning.

What is Unsupervised Learning?

Unsupervised learning is about finding hidden structure in data without being told what to look for. There are no labels, no “correct answers” - just raw data, and the model’s job is to discover patterns.

flowchart LR
  A["Unlabeled Data
(no correct answers)"] --> B["Unsupervised
Algorithm"]
  B --> C["Discovered Structure
(clusters, patterns,
reduced dimensions)"]
  style B fill:#e8f4fd,stroke:#1a5276,color:#1a5276
  style C fill:#e8f8f0,stroke:#1a5c38,color:#1a5c38

Supervised vs Unsupervised

Why Unsupervised Learning Matters

In the real world, most data is unlabeled. Labeling data is expensive, slow, and sometimes impossible:

• A hospital has millions of patient records but no one has tagged each record with disease labels
• An e-commerce site has billions of clickstreams but hasn’t categorized every user
• A security system has network logs but doesn’t know which are attacks

Unsupervised learning lets you extract value from raw, unlabeled data.

Types of Unsupervised Learning

flowchart TD
  UL["Unsupervised Learning"] --> CL["Clustering"]
  UL --> DR["Dimensionality Reduction"]
  UL --> AD["Anomaly Detection"]
  UL --> AR["Association Rules"]
  CL --> KM["K-Means"]
  CL --> HC["Hierarchical"]
  CL --> DB["DBSCAN"]
  DR --> PCA["PCA"]
  DR --> TSNE["t-SNE"]
  DR --> UMAP["UMAP"]
  AD --> IF["Isolation Forest"]
  AD --> AE["Autoencoders"]
  AR --> APR["Apriori"]
  style UL fill:#f3e8fd,stroke:#5b1a7a,color:#5b1a7a
  style CL fill:#e8f8f0,stroke:#1a5c38,color:#1a5c38
  style DR fill:#e8f4fd,stroke:#1a5276,color:#1a5276
  style AD fill:#fdf3e8,stroke:#7a4a1a,color:#7a4a1a
  style AR fill:#fde8e8,stroke:#7a1a1a,color:#7a1a1a

Real-World Applications

Clustering: Grouping Similar Data

Clustering is the most common unsupervised task. The goal: group data points so that points in the same group (cluster) are more similar to each other than to points in other groups.

K-Means Clustering

The most popular clustering algorithm. Here’s how it works:

flowchart TD
  A["1. Choose K
(number of clusters)"] --> B["2. Randomly place K
cluster centers (centroids)"]
  B --> C["3. Assign each point
to nearest centroid"]
  C --> D["4. Move each centroid
to mean of its points"]
  D --> E{"Centroids
moved?"}
  E -->|"Yes"| C
  E -->|"No"| F["Done! Clusters found"]
  style A fill:#e8f4fd,stroke:#1a5276,color:#1a5276
  style C fill:#fdf3e8,stroke:#7a4a1a,color:#7a4a1a
  style D fill:#fdf3e8,stroke:#7a4a1a,color:#7a4a1a
  style F fill:#e8f8f0,stroke:#1a5c38,color:#1a5c38

Example: Customer Segmentation

An e-commerce company has customer data with no predefined groups:

After running K-Means with K=3:

No one told the algorithm these groups existed - it discovered them from the data.

Choosing K: The Elbow Method

How do you know the right number of clusters? Plot the inertia (sum of squared distances to nearest centroid) for different values of K:

The “elbow” is at K=3 - after that, adding more clusters gives diminishing returns.

K-Means Limitations

Hierarchical Clustering

Unlike K-Means, hierarchical clustering doesn’t require you to specify K upfront. It builds a tree (dendrogram) of clusters.

flowchart TD
  A["Agglomerative
(Bottom-Up)"] --> A1["Start: each point is
its own cluster"]
  A1 --> A2["Merge the two closest
clusters"]
  A2 --> A3["Repeat until one
cluster remains"]
  B["Divisive
(Top-Down)"] --> B1["Start: all points in
one cluster"]
  B1 --> B2["Split the least
coherent cluster"]
  B2 --> B3["Repeat until each point
is its own cluster"]
  style A fill:#e8f8f0,stroke:#1a5c38,color:#1a5c38
  style B fill:#e8f4fd,stroke:#1a5276,color:#1a5276

You choose the number of clusters by “cutting” the dendrogram at a desired height.

DBSCAN: Density-Based Clustering

DBSCAN (Density-Based Spatial Clustering of Applications with Noise) finds clusters of arbitrary shape and automatically identifies outliers.

How DBSCAN Classifies Points

K-Means vs DBSCAN

Dimensionality Reduction

Real-world datasets often have dozens or hundreds of features. Dimensionality reduction compresses data to fewer dimensions while preserving important information.

Why Reduce Dimensions?

• Visualization: Plot high-dimensional data in 2D/3D
• Speed: Fewer features → faster training
• Noise reduction: Remove irrelevant features
• Avoid curse of dimensionality: Models perform worse with too many features

PCA (Principal Component Analysis)

PCA finds the directions of maximum variance in the data and projects onto those directions.

flowchart LR
  A["Original Data
100 features"] --> B["PCA"]
  B --> C["Reduced Data
2-3 features"]
  B --> D["Variance Explained:
PC1: 45%
PC2: 25%
PC3: 15%"]
  style B fill:#e8f4fd,stroke:#1a5276,color:#1a5276
  style C fill:#e8f8f0,stroke:#1a5c38,color:#1a5c38

PCA Example: Student Performance

Original features (6 dimensions):

PCA might discover:

• PC1 = “STEM aptitude” (positive correlation with Math, Physics, Chemistry)
• PC2 = “Humanities aptitude” (positive correlation with English, History, Art)

6 dimensions compressed to 2 - and the essential structure is preserved.

Anomaly Detection

Find data points that are significantly different from the rest. Critical for:

• Fraud detection: Unusual credit card transactions
• Network security: Abnormal traffic patterns
• Manufacturing: Defective products on assembly lines
• Health monitoring: Abnormal vital signs

How It Works

Example: Credit Card Fraud

The fraudulent transaction has an extremely high anomaly score - different from normal patterns in amount, time, and location.

Evaluating Unsupervised Models

Without labels, evaluation is harder. Common approaches:

Supervised vs Unsupervised: When to Use Which

Summary

Series Complete!

flowchart LR
  A["✅ Intro to ML"] --> B["✅ Supervised Learning"]
  B --> C["✅ Regression"]
  C --> D["✅ Classification"]
  D --> E["✅ Unsupervised Learning"]
  style A fill:#e8f8f0,stroke:#1a5c38,color:#1a5c38
  style B fill:#e8f8f0,stroke:#1a5c38,color:#1a5c38
  style C fill:#e8f8f0,stroke:#1a5c38,color:#1a5c38
  style D fill:#e8f8f0,stroke:#1a5c38,color:#1a5c38
  style E fill:#e8f8f0,stroke:#1a5c38,color:#1a5c38

You now have a solid foundation in the core concepts of machine learning:

1. What ML is and how it differs from traditional programming
2. Supervised learning - learning from labeled data
3. Regression - predicting continuous values with linear models
4. Classification - predicting categories with logistic regression and beyond
5. Unsupervised learning - discovering hidden patterns without labels

From here, you can dive deeper into specific algorithms, explore deep learning, or start building real models with Python and scikit-learn. The fundamentals you’ve built here will carry you through all of it.

Happy learning.