CT032-3-3-FAI& Further Artificial Intelligence

UNSUPERVISED LEARNING

Module Code & Module Title Slide Title SLIDE 1

 TOPIC LEARNING OUTCOMES
At the end of this topic, you should be able to:
1. Explain the fundamental concepts of unsupervised learning
2. Distinguish between key clustering algorithms
3. Understand dimensionality reduction techniques
4. Apply basic clustering algorithms to real-world datasets
5. Evaluate the quality of clustering results

Module Code & Module Title Slide Title SLIDE 2

 CONTENTS & STRUCTURE
1. Fundamental concepts of unsupervised learning
2. Key clustering algorithms
3. Dimensionality reduction techniques
4. Quality of clustering results

Module Code & Module Title Slide Title SLIDE 3

 What is Unsupervised Learning?

Core Concepts of Unsupervised Learning:

1. Unlabelled Data
2. Pattern Recognition
3. Data Distribution Understanding
4. Dimensionality
5. Iterative Nature
6. Evaluation Challenges

Image credits: google.com

Module Code & Module Title Slide Title SLIDE 4
 What is Unsupervised Learning?...cont.

Unlabeled Data
• No labels are given to the learning algorithm, leaving it on its
own to find structure in its input.
• Unsupervised machine learning cannot be directly applied to a
regression because it is unknown what the output values could be,
therefore making it impossible to train the algorithm how you
normally would.

Image credits: google.com

Module Code & Module Title Slide Title SLIDE 5
 What is Unsupervised Learning?...cont.

Pattern Recognition
• This could be finding:
1. Natural groupings of similar items
2. Hidden relationships between variables
3. Common features or characteristics
4. Abnormal or unusual patterns

Image credits: google.com

Module Code & Module Title Slide Title SLIDE 6
 What is Unsupervised Learning?...cont.

Data Distribution Understanding

• Unsupervised learning helps understand how data is distributed
and organized in the feature space.

• This includes:
1. Density estimation: Understanding where data points are
concentrated
2. Feature relationships: How different variables correlate or
interact
3. Data structure: The inherent organization or hierarchy in the
data
Image credits: google.com
Module Code & Module Title Slide Title SLIDE 7
 What is Unsupervised Learning?...cont.

Dimensionality
• Many unsupervised learning tasks involve working with high-
dimensional data where patterns are not immediately obvious.

• The algorithms help by:

1. Reducing the dimensions while retaining important
information
2. Finding the most important features
3. Creating more compact representations of the data

Image credits: google.com

Module Code & Module Title Slide Title SLIDE 8
 What is Unsupervised Learning?...cont.

Iterative Nature
• K-means is an unsupervised clustering algorithm designed to
partition unlabelled data into a certain number (which is the “ K”) of
distinct groupings.

Image credits: towardsdatascience.com

Module Code & Module Title Slide Title SLIDE 9
 What is Unsupervised Learning?...cont.

Iterative Nature

Image credits: towardsdatascience.com

Module Code & Module Title Slide Title SLIDE 10
 What is Unsupervised Learning?...cont.

Evaluation Challenges

• Unlike supervised learning, there's often no clear "right answer" to

validate against.

• Success is typically measured by:

1. Internal metrics (like cluster cohesion)
2. Domain expertise validation
3. Practical usefulness of the discovered patterns

Image credits: google.com

Module Code & Module Title Slide Title SLIDE 11
 Quick Review Question

Why is unsupervised learning called "unsupervised" and what is its

main goal?

Image credits: google.com

Module Code & Module Title Slide Title SLIDE 12
 Key Clustering Algorithms

K-means Clustering
• Example: Segmentation in a gym:
– Input: Member data (age, visit frequency, class preferences)
• Process: Divides into K groups (e.g., K=3)
• Results:
– Cluster 1: Young, frequent visitors, prefer group classes
– Cluster 2: Middle-aged, moderate visits, prefer machines
– Cluster 3: Seniors, regular visits, prefer morning sessions

Image credits: google.com

Module Code & Module Title Slide Title SLIDE 13
 Key Clustering Algorithms ...cont.

Hierarchical Clustering
• Example: Organizing a company's product catalog:
• Top Level:
• Electronics, Clothing, Home goods
• Second Level:
• Electronics → Phones, Laptops, Accessories
• Clothing → Men's, Women's, Children's
• Third Level:
• Phones → Budget, Mid-range, Premium Creates a tree-like
structure showing relationships at different levels.

Image credits: google.com

Module Code & Module Title Slide Title SLIDE 14
 Key Clustering Algorithms ...cont.

DBSCAN (Density-Based)
• Example: Analyzing city locations:
• Core Points: Dense areas (shopping districts)
• Border Points: Edge of neighborhoods
• Noise: Isolated locations

• Finds irregularly shaped clusters like:

1. Shopping districts of any shape
2. Connected neighborhoods
3. Identifies outliers (isolated stores)

Image credits: google.com

Module Code & Module Title Slide Title SLIDE 15
 Key Clustering Algorithms ...cont.

Gaussian Mixture Models

• Example: Analyzing customer spending:
• Overlapping groups based on probability
• Can identify:
• Budget shoppers ($0-50 range)
• Mid-range ($30-150 range)
• Luxury ($100+ range) Allows for fuzzy boundaries where
customers might belong partially to multiple groups.

Image credits: google.com

Module Code & Module Title Slide Title SLIDE 16
 Quick Review Question

What's the key difference between K-means and DBSCAN

clustering?

Image credits: google.com

Module Code & Module Title Slide Title SLIDE 17

 Dimensionality Reduction Techniques

Principal Component Analysis (PCA)

• Example: Analyzing smartphone features
– Original 6 dimensions:
• Screen size, Battery life, Camera quality, Processor speed, RAM,
Storage

• Reduced to 2 key components:

• PC1: "Performance" (combines processor, RAM, storage)
• PC2: "User Experience" (combines screen, battery, camera)

• This simplifies analysis while maintaining most important

information. Image credits: google.com
Module Code & Module Title Slide Title SLIDE 18
 Dimensionality Reduction Techniques ...cont.

t-SNE (t-Distributed Stochastic Neighbor Embedding)

• Example: Visualizing customer behaviors
• Original dimensions:
• Purchase history, Browsing patterns, Time spent on site, Cart
abandonment, Return history, Review ratings

• Reduces to 2D visualization showing:

• Clear clusters of similar customers
• Outlier detection

• Relationship patterns Especially good for visualizing complex

patterns. Image credits: google.com
Module Code & Module Title Slide Title SLIDE 19
 Dimensionality Reduction Techniques ...cont.

Autoencoders
• Example: Image compression
• Input: 1000x1000 pixel image
• Process:
• Encoder compresses to key features
• Decoder reconstructs from compressed form

• Learns efficient representation Output: Same image reconstructed

from much less data

Image credits: google.com

Module Code & Module Title Slide Title SLIDE 20
 Dimensionality Reduction Techniques ...cont.

LDA (Linear Discriminant Analysis)

• Example: Text document analysis
• Original: Thousands of words per document
• Reduced to:
• Key topics
• Main themes

• Document similarities Helps in document classification and topic

modeling.

Image credits: google.com

Module Code & Module Title Slide Title SLIDE 21
 Quick Review Question

When would you choose t-SNE over PCA for

dimensionality reduction?

Image credits: google.com

Module Code & Module Title Slide Title SLIDE 22

 Quality of Clustering Results

Silhouette Score
• Example: Analyzing customer segments
• Score range: -1 to 1
• Higher is better
• Example scores:
• 0.8: Clear separation between young, middle-aged, and senior
shoppers
• 0.3: Some overlap in shopping patterns between age groups
• -0.1: Poor clustering, age groups not meaningfully separated

Image credits: google.com

Module Code & Module Title Slide Title SLIDE 23
 Quality of Clustering Results...cont.

Intra-Cluster Distance
• Example: Restaurant groupings by cuisine and price
• Good clustering:
• All Italian fine dining restaurants tightly grouped
• All fast food chains closely clustered Poor clustering:
• Mix of expensive and budget restaurants in same cluster
• Different cuisines randomly mixed together

Image credits: google.com

Module Code & Module Title Slide Title SLIDE 24
 Quality of Clustering Results...cont.

Inter-Cluster Distance
• Example: Product categories in e-commerce
• Good separation:
• Electronics cluster clearly distinct from clothing cluster
• Beauty products well-separated from tools Poor separation:
• Sports equipment overlapping with casual wear
• Home decor mixing with kitchen appliances

Image credits: google.com

Module Code & Module Title Slide Title SLIDE 25
 Quality of Clustering Results...cont.

Stability Analysis
• Example: Customer behavior clustering
• Stable results:
• Similar clusters appear even with different data samples
• Patterns remain consistent over time Unstable results:
• Clusters change dramatically with small data changes
• No consistent patterns emerge

Image credits: google.com

Module Code & Module Title Slide Title SLIDE 26
 Quick Review Question

What does a high silhouette score (close to 1) tell you

about your clustering results?

Image credits: google.com

Module Code & Module Title Slide Title SLIDE 27

 Summary / Recap of Main Points

• Fundamental Concepts of Unsupervised Learning

• Works with unlabeled data – no predefined correct answers
• Focuses on finding hidden patterns and structures
• Aims to understand data distribution and relationships
• Handles high-dimensional data through reduction
• Uses iterative processes to improve results
• Requires careful evaluation due to lack of ground truth

Module Code & Module Title Slide Title SLIDE 28

 Summary / Recap of Main Points…cont.

• Key Clustering Algorithms

• K-means: Simple, fast, requires predefined number of clusters
• Hierarchical Clustering: Creates tree-like structure, good for
nested relationships
• DBSCAN: Handles irregular shapes, identifies outliers
automatically
• Gaussian Mixture Models: Allows probabilistic cluster
membership Each suitable for different types of data and
objectives

Module Code & Module Title Slide Title SLIDE 29

 Summary / Recap of Main Points…cont.

• Dimensionality Reduction Techniques

• PCA: Linear reduction, preserves maximum variance
• t-SNE: Good for visualization, preserves local relationships
• Autoencoders: Non-linear reduction using neural networks
• LDA: Supervised reduction, optimizes class separation Choose
based on data type and visualization needs

Module Code & Module Title Slide Title SLIDE 30

 Summary / Recap of Main Points…cont.

• Quality of Clustering Results

• Silhouette Score: Measures cluster cohesion and separation
• Intra-Cluster Distance: Evaluates similarity within clusters
• Inter-Cluster Distance: Assesses separation between clusters
• Stability Analysis: Tests robustness of clustering results
Essential for validating clustering effectiveness

Module Code & Module Title Slide Title SLIDE 31

 What To Expect Next Week

In Class Preparation for Class

• Unsupervised Learning • Clustering & Classification

Module Code & Module Title Slide Title SLIDE 32