Unit 3 — Machine Learning | Study Notes

UNIT 3
Machine Learning
Fundamentals, Types & Applications

Topics Covered
Q11. Definition & Features of Machine Learning
Q12. Concept & Importance of Machine Learning
Q13. Relationship between AI and Machine Learning
Q14. Comparison of AI and Machine Learning
Q15. Types of Machine Learning
Q16 & Q17. Supervised Learning with Examples
Q18. Unsupervised Learning and Applications

Page 1 of 19
Unit 3 — Machine Learning | Study Notes

Q11. Definition & Important Features of Machine Learning

Definition
Machine Learning (ML) is a subset of Artificial Intelligence (AI) that enables computer systems to
automatically learn and improve from experience without being explicitly programmed for every
task. ML algorithms build mathematical models from sample data, known as training data, and use
those models to make predictions or decisions on new, unseen data.

Key Definition (Arthur Samuel, 1959): "Machine Learning is the field of study that gives
computers the ability to learn without being explicitly programmed."

In modern terms, Machine Learning is the process by which a computer system identifies patterns
in large datasets, learns from those patterns, and uses that knowledge to perform tasks intelligently
— improving its performance over time with more data.

Important Features of Machine Learning

1. Automated Learning from Data

ML systems acquire knowledge directly from data without the need for human programmers to
define every rule explicitly. The model updates its understanding as it processes more training
examples.

2. Data-Driven Approach
The performance and accuracy of an ML model are entirely dependent on the quality, quantity, and
diversity of the data used to train it. More relevant data leads to better predictions.

3. Pattern Recognition
ML excels at detecting hidden patterns, trends, and relationships within large, complex, and high-
dimensional datasets — tasks that would be practically impossible for humans to do manually.

4. Generalization
A well-trained ML model does not simply memorize training data; it generalizes from it — applying
learned patterns to new, unseen inputs accurately. This property is what makes ML practically
useful.

5. Adaptability
ML models can adapt dynamically to changing data distributions. When exposed to new
information, models can be retrained to maintain accuracy, making them suitable for evolving real-
world environments.

Page 2 of 19
Unit 3 — Machine Learning | Study Notes

6. Prediction and Decision-Making

ML models generate probabilistic predictions and support intelligent, data-driven decision-making
across domains ranging from healthcare diagnosis to stock price forecasting.

7. Minimal Human Intervention

Once trained and deployed, ML models operate with minimal human oversight, automating tasks
that previously required continuous manual effort.

8. Scalability
ML algorithms scale efficiently with increasing data volumes. As more data becomes available,
models can be retrained to deliver higher accuracy without fundamentally changing their
architecture.

9. Self-Improvement
Through iterative training cycles and feedback loops, ML systems improve their accuracy and
reliability over time — a property that distinguishes ML from conventional rule-based software.

10. Wide Applicability

ML techniques are domain-agnostic and have been successfully applied in healthcare, finance,
transportation, agriculture, natural language processing, computer vision, and many other fields.

Page 3 of 19
Unit 3 — Machine Learning | Study Notes

Q12. Concept & Importance of Machine Learning in Modern

Technology

Concept of Machine Learning

Machine Learning is based on the idea that systems can learn from data, identify patterns, and
make decisions with minimal human intervention. The core concept revolves around three
fundamental components:

• Data: Data is the raw material fed into the ML system. It includes labeled or unlabeled
examples from which the model learns.
• Algorithm: An Algorithm is the mathematical procedure that processes data and builds a
predictive model. Common algorithms include decision trees, neural networks, and support
vector machines.
• Model: The Model is the output of training — a mathematical representation that maps
inputs to outputs. It is then deployed to make predictions on real-world data.

The ML Workflow: Collect Data → Preprocess Data → Choose Algorithm → Train Model →
Evaluate Model → Deploy Model → Monitor & Retrain

The learning process can be understood as an optimization problem: given a dataset, the algorithm
adjusts its internal parameters to minimize a loss function — a measure of how far the model's
predictions deviate from the actual outcomes. Over thousands or millions of iterations, the model
converges towards optimal parameters that yield accurate predictions.

Importance of Machine Learning in Modern Technology

1. Automation of Repetitive Tasks

ML automates time-consuming, repetitive tasks such as data entry, document classification, and
quality control in manufacturing, freeing human workers for more creative and strategic
responsibilities.

2. Intelligent Personalization
Platforms like Netflix, Spotify, Amazon, and YouTube use ML to analyze user behavior and deliver
personalized recommendations, dramatically improving user engagement and satisfaction.

3. Healthcare and Medical Diagnostics

ML models analyze medical images (X-rays, MRIs) with accuracy comparable to expert
radiologists. They also accelerate drug discovery, predict patient deterioration, and personalize
treatment plans.

Page 4 of 19
Unit 3 — Machine Learning | Study Notes

4. Natural Language Processing (NLP)

ML powers virtual assistants (Siri, Alexa, Google Assistant), machine translation (Google
Translate), sentiment analysis, and large language models like ChatGPT — enabling machines to
understand and generate human language.

5. Fraud Detection and Cybersecurity

Banks and financial institutions use ML to detect fraudulent transactions in real time by identifying
anomalous patterns that deviate from a customer's typical behavior, preventing billions in losses.

6. Autonomous Vehicles
Self-driving cars rely on ML algorithms to process sensor data, recognize objects, predict the
behavior of pedestrians and other vehicles, and make split-second driving decisions safely.

7. Predictive Analytics
Businesses use ML for demand forecasting, predictive maintenance of industrial equipment, supply
chain optimization, and financial risk modeling — enabling proactive rather than reactive decision-
making.

8. Scientific Discovery
ML accelerates research in genomics, climate science, materials science, and astronomy by
identifying patterns in massive experimental datasets that would take human researchers decades
to analyze manually.

Page 5 of 19
Unit 3 — Machine Learning | Study Notes

Q13. Relationship Between Artificial Intelligence and Machine

Learning

Overview
Artificial Intelligence (AI) and Machine Learning (ML) are closely related fields, but they are not
synonymous. Understanding their relationship requires examining their scope, goals, and how they
interact with each other.

Relationship Summary: Artificial Intelligence is the broad discipline of creating intelligent

machines. Machine Learning is one of the most powerful methodologies used to achieve
Artificial Intelligence. All Machine Learning is AI, but not all AI is Machine Learning.

Artificial Intelligence — The Broader Discipline

Artificial Intelligence is the overarching field of computer science focused on building systems that
can perform tasks that typically require human intelligence. These tasks include reasoning,
problem-solving, understanding natural language, recognizing patterns, planning, and learning.
AI encompasses multiple approaches including rule-based expert systems, search algorithms, logic
programming, robotics, computer vision, and machine learning. Early AI systems (1950s–1980s)
were primarily rule-based: programmers manually encoded knowledge as IF-THEN rules, enabling
computers to simulate expert decision-making in narrow domains.

Machine Learning — The Subset That Learns

Machine Learning emerged as a subset of AI in response to the limitations of rule-based systems: it
is extremely difficult for humans to anticipate and encode every possible scenario, especially in
complex, high-dimensional domains. ML addressed this by allowing machines to learn rules directly
from data.
Instead of programming explicit rules, ML practitioners provide algorithms with large datasets and
allow the algorithm to discover the rules autonomously through statistical analysis and optimization.

Hierarchical Relationship

Artificial Intelligence (Broadest Scope)

Machine Learning (Subset of AI — Learns from Data)

Deep Learning (Subset of ML — Uses Neural Networks)

Page 6 of 19
Unit 3 — Machine Learning | Study Notes

How ML Enables AI
• Learning Capability: ML provides AI systems the ability to improve performance on tasks
over time without being reprogrammed.
• Generalization: ML allows AI to generalize from training examples and handle situations
not explicitly anticipated by programmers.
• Scalability: ML-based AI systems can scale to handle complex, high-dimensional data
(images, speech, text) far beyond the capacity of rule-based systems.
• Feature Discovery: ML lets AI discover features and patterns automatically from raw data,
eliminating the need for manual feature engineering.

Page 7 of 19
Unit 3 — Machine Learning | Study Notes

Q14. Comparison of Artificial Intelligence and Machine Learning

Comparison Table

Aspect Artificial Intelligence (AI) Machine Learning (ML)

Definition The broad science of making A subset of AI that enables

machines simulate human machines to learn from data
intelligence automatically
Scope Broader — includes ML, Narrower — focused
robotics, expert systems, NLP, specifically on learning from
planning data
Goal To build intelligent machines To build models that improve
that can perform any cognitive performance through data-
task driven learning
Approach Can use rules, logic, search, Always uses statistical models
optimization, or learning and algorithms trained on data
Human Involvement Varies — rule-based AI Requires data preparation;
requires heavy human input of rules are discovered
rules automatically
Adaptability Rule-based AI is rigid; may not Naturally adaptive —
adapt without reprogramming performance improves with
more data
Decision Making Can use fixed logic or learned Always based on patterns
models learned from training data
Examples Chess engines (Deep Blue), Spam filters, Recommendation
Expert Systems, Robotics engines, Image recognition
Dependency AI does not necessarily depend ML always falls under and
on ML depends on AI as its parent
field
Data Requirement Not always data-dependent Highly data-dependent —
(e.g., rule-based systems) performance scales with data
quality/quantity

Examples Illustrating the Difference

Example 1 — Email Spam Detection

• AI Approach: An AI rule-based system might use fixed rules: block emails containing the
word 'free money' or from unknown senders.

Page 8 of 19
Unit 3 — Machine Learning | Study Notes

• ML Approach: An ML spam filter analyzes thousands of labeled emails (spam/not spam),

learns linguistic patterns, sender behaviors, and structural features, then predicts whether
new emails are spam — adapting as spammers evolve.

Example 2 — Playing Chess

• AI Approach: IBM's Deep Blue defeated Garry Kasparov using a rule-based AI approach: it
searched millions of possible moves using programmed heuristics and evaluation functions
— no learning from data.
• ML Approach: AlphaZero (Google DeepMind) used ML (reinforcement learning) to learn
chess from scratch by playing against itself millions of times, discovering strategies that
surpassed all human and traditional AI systems.

Example 3 — Medical Diagnosis

• AI Approach: A traditional AI expert system uses rules encoded by medical specialists: IF
fever > 38.5 AND cough THEN consider influenza.
• ML Approach: An ML diagnostic model is trained on millions of patient records, lab results,
and confirmed diagnoses. It learns complex, multi-factor patterns that no human expert
could fully articulate, often achieving higher diagnostic accuracy.

Page 9 of 19
Unit 3 — Machine Learning | Study Notes

Q15. Classification of Different Types of Machine Learning

Machine Learning is broadly classified into four main types based on the nature of learning, the type
of feedback provided to the algorithm, and the structure of the training data:

1. Supervised Learning

Definition: The algorithm is trained on a labeled dataset — every input data point is paired with a
corresponding correct output (label). The model learns to map inputs to outputs by minimizing
prediction errors.

In supervised learning, the learning process is analogous to a student learning under a teacher's
supervision: the teacher provides correct answers, the student practices, and errors are corrected
until the student masters the subject.

• Sub-type 1: Classification — predicting a discrete category label (e.g., spam/not spam,

disease/no disease)
• Sub-type 2: Regression — predicting a continuous numerical value (e.g., house price,
temperature, stock value)

Examples: Email spam detection, image classification, medical diagnosis, credit scoring, stock price
prediction.

2. Unsupervised Learning

Definition: The algorithm is trained on an unlabeled dataset — no correct outputs are provided.
The model must discover hidden structure, patterns, or groupings in the data on its own.

Unsupervised learning is like exploring an unknown territory without a map — the system must
make sense of its environment autonomously, finding natural groupings and structures.

• Sub-type 1: Clustering — grouping similar data points together (e.g., customer

segmentation, document clustering)
• Sub-type 2: Dimensionality Reduction — compressing data while retaining important
information (e.g., PCA, t-SNE)
• Sub-type 3: Association Rule Mining — discovering relationships between variables in large
datasets (e.g., market basket analysis)

Examples: Customer segmentation, anomaly detection, topic modeling, gene expression analysis,
recommendation systems.

Page 10 of 19
Unit 3 — Machine Learning | Study Notes

3. Semi-Supervised Learning

Definition: The algorithm is trained on a small amount of labeled data combined with a large
amount of unlabeled data. This approach bridges supervised and unsupervised learning, making
it practical when labeling data is expensive or time-consuming.

Semi-supervised learning is highly valuable in real-world scenarios where obtaining labeled data
requires significant manual effort (e.g., medical image annotation by radiologists). The model uses
the labeled examples as anchors and leverages the structure of unlabeled data to improve
generalization.

Examples: Web content classification, speech recognition, protein structure prediction, medical
imaging when labeled scans are scarce.

4. Reinforcement Learning

Definition: An agent learns to make sequential decisions by interacting with an environment. It

receives rewards for correct actions and penalties for incorrect ones, gradually learning a policy
that maximizes cumulative reward over time.

Reinforcement learning mimics how humans and animals learn through trial and error: a child
learns to walk by trying, falling, and adjusting — the reward is successful locomotion. The agent has
no labeled dataset; it generates its own experience through interaction.

• Key Component 1: Agent — the entity that takes actions (e.g., a game-playing AI, a robot)
• Key Component 2: Environment — the external system the agent interacts with (e.g., a
game board, physical world)
• Key Component 3: State — the current situation of the agent in the environment
• Key Component 4: Action — a decision the agent makes at each time step
• Key Component 5: Reward — feedback signal indicating the quality of the action taken
• Key Component 6: Policy — the strategy the agent learns; maps states to actions to
maximize reward

Examples: Game playing (AlphaGo, AlphaZero), robotics, autonomous driving, personalized

education platforms, financial trading algorithms.

Page 11 of 19
Unit 3 — Machine Learning | Study Notes

Q16 & Q17. Supervised Learning with Suitable Examples

Definition and Concept

Supervised Learning is a type of Machine Learning in which the algorithm is trained on a labeled
dataset — a collection of input-output pairs where every input (feature vector) is associated with a
known, correct output (label or target value). The term 'supervised' refers to the guidance provided
by these labels, which act like a supervisor correcting the model during training.

Formal Definition: Given a training set D = {(x1, y1), (x2, y2), ..., (xn, yn)}, where xi represents
the input features and yi represents the corresponding output label, a supervised learning
algorithm learns a function f such that f(xi) ≈ yi for all training examples, and generalizes
accurately to new unseen inputs.

How Supervised Learning Works

1. Data Collection: Gather a large, representative labeled dataset.

2. Data Preprocessing: Clean, normalize, encode, and split data into training, validation, and
test sets.
3. Model Selection: Choose an appropriate algorithm (e.g., Linear Regression, SVM, Decision
Tree, Neural Network).
4. Model Training: Feed the training data to the algorithm. The model adjusts its parameters to
minimize the difference between predicted and actual outputs (loss function).
5. Model Evaluation: Assess the trained model on validation/test data using metrics like
accuracy, precision, recall, F1-score, or RMSE.
6. Prediction: Deploy the model to predict outputs for new, unseen inputs.

Types of Supervised Learning

A) Classification
Classification predicts a discrete category label. The output is one of a finite set of predefined
classes. The model learns decision boundaries that separate the feature space into distinct class
regions.

• Binary Classification: Two possible outputs (e.g., spam/not spam, tumor/benign, pass/fail)
• Multi-class Classification: More than two possible outputs (e.g., handwritten digit recognition:
0–9; species classification)
• Multi-label Classification: Each input can belong to multiple classes simultaneously (e.g.,
tagging a news article with topics: politics, economy, health)

Page 12 of 19
Unit 3 — Machine Learning | Study Notes

B) Regression
Regression predicts a continuous numerical value. The model learns a mapping from input features
to a real-valued output. The goal is to minimize the average error between predicted and actual
values.

• Linear Regression: Models a linear relationship between inputs and output

• Polynomial Regression: Models non-linear relationships using polynomial features
• Multiple Regression: Uses multiple input features to predict a single output value

Common Supervised Learning Algorithms

Algorithm Type Best Used For

Linear Regression Regression Predicting continuous values

(prices, temperatures)
Logistic Regression Classification Binary classification problems
Decision Trees Both Interpretable models for
classification and regression
Random Forests Both High-accuracy ensemble
learning on tabular data
Support Vector Machines Classification High-dimensional data, text and
(SVM) image classification
K-Nearest Neighbors (KNN) Both Simple classification and
regression tasks
Neural Networks Both Complex patterns in images,
text, speech, and video
Naive Bayes Classification Text classification, spam
detection, sentiment analysis

Detailed Examples of Supervised Learning

Example 1 — Email Spam Detection (Classification)

Task: Classify incoming emails as either SPAM or NOT SPAM (binary classification).

• Input Features: Email subject line words, sender address, presence of hyperlinks, email
body keywords (free, win, lottery), email length, HTML content
• Labels: 'Spam' (1) or 'Not Spam' (0) — provided by human annotators reviewing thousands
of emails

Page 13 of 19
Unit 3 — Machine Learning | Study Notes

• Algorithm Used: Naive Bayes Classifier or Logistic Regression

• Training: Model learns which word combinations and structural features are statistically
associated with spam emails
• Outcome: A deployed spam filter that correctly classifies new emails in real time with over
99% accuracy on modern systems

Example 2 — House Price Prediction (Regression)

Task: Predict the sale price of a house based on its characteristics.

• Input Features: Number of bedrooms, area in square feet, location (zip code), age of
property, number of bathrooms, garage capacity, proximity to schools
• Labels: Actual sale prices (in USD) from historical real estate transaction records
• Algorithm Used: Multiple Linear Regression or Gradient Boosted Trees (XGBoost)
• Training: Model learns the weighted contribution of each feature to the final price
• Outcome: A model that predicts house prices within a small margin of error, used by real
estate platforms like Zillow (Zestimate)

Example 3 — Medical Image Classification (Classification)

Task: Detect diabetic retinopathy from retinal fundus photographs.

• Input Features: Pixel values of high-resolution retinal images (2D arrays of RGB values)
• Labels: Expert ophthalmologist diagnoses categorized into 5 severity levels (No DR, Mild,
Moderate, Severe, Proliferative DR)
• Algorithm Used: Convolutional Neural Network (CNN) — a deep learning architecture
• Training: Model trained on over 88,000 labeled retinal images
• Outcome: Google's DiabeticRetinopathy model achieved accuracy matching or exceeding
expert ophthalmologists, enabling early screening in underserved regions

Example 4 — Credit Score Prediction (Classification + Regression)

Task: Predict whether a loan applicant will default on their loan.

• Input Features: Income, employment status, existing debt, credit history length, number of
late payments, loan amount requested
• Labels: 'Default' or 'No Default' from historical loan repayment records
• Algorithm Used: Random Forest Classifier or Gradient Boosting
• Outcome: Banks and fintech companies use these models to approve or reject loan
applications instantly, reducing default risk significantly

Page 14 of 19
Unit 3 — Machine Learning | Study Notes

Advantages of Supervised Learning

• High accuracy when sufficient labeled data is available
• Clear performance measurement using well-defined evaluation metrics
• Wide range of proven algorithms available for different problem types
• Models are interpretable and explainable (especially Decision Trees and Linear Regression)

Limitations of Supervised Learning

• Requires large amounts of labeled data, which is expensive and time-consuming to prepare
• Poor performance when training data does not accurately represent real-world data
(distribution shift)
• Susceptible to overfitting — memorizing training data rather than generalizing
• Cannot discover insights beyond what is represented in the labeled training data

Page 15 of 19
Unit 3 — Machine Learning | Study Notes

Q18. Unsupervised Learning and Its Applications

Definition
Unsupervised Learning is a type of Machine Learning in which the algorithm is trained on data that
has no predefined labels or correct output values. The model is not given explicit guidance about
what to look for; instead, it must autonomously discover hidden patterns, natural groupings,
structures, or relationships within the raw data.

Key Distinction: In Supervised Learning, the model is told what the correct answer is. In
Unsupervised Learning, the model must figure out the structure of data entirely on its own —
there is no 'teacher.'

Unsupervised learning is particularly valuable when labeled data is unavailable, expensive to

obtain, or when the goal is to explore the data to generate new hypotheses rather than validate
existing ones. It is the closest form of ML to how humans naturally learn — observing patterns in
the environment without explicit instruction.

Types of Unsupervised Learning

1. Clustering
Clustering partitions a dataset into groups (clusters) such that data points within the same cluster
are more similar to each other than to points in other clusters. The algorithm determines both the
number of clusters and their boundaries (in most algorithms) from the data itself.

• K-Means Clustering: Partitions data into K clusters by iteratively assigning points to the
nearest cluster centroid and updating centroids. Fast and scalable but requires specifying K
in advance.
• Hierarchical Clustering: Builds a tree-like hierarchy of clusters (dendrogram) by successively
merging or splitting clusters based on similarity. Does not require specifying K in advance.
• DBSCAN (Density-Based Spatial Clustering): Groups points that are closely packed
together, marking points in low-density regions as outliers. Excellent for discovering
arbitrarily-shaped clusters and detecting anomalies.
• Gaussian Mixture Models (GMM): Probabilistic model that assumes data is generated from
a mixture of Gaussian distributions, allowing soft cluster assignments.

2. Dimensionality Reduction
Dimensionality reduction transforms high-dimensional data into a lower-dimensional representation
while preserving as much meaningful information as possible. This is critical for visualization, noise
reduction, and improving computational efficiency.

Page 16 of 19
Unit 3 — Machine Learning | Study Notes

• Principal Component Analysis (PCA): Linear transformation that projects data onto a new
set of orthogonal axes (principal components) that capture maximum variance.
• t-SNE (t-distributed Stochastic Neighbor Embedding): Non-linear technique primarily used
for 2D/3D visualization of high-dimensional data, preserving local relationships between
points.
• Autoencoders: Deep learning-based dimensionality reduction where a neural network learns
a compressed representation (encoding) and then reconstructs the original input (decoding).

3. Association Rule Learning

Association rule learning discovers interesting relationships and co-occurrence patterns between
variables in large transactional databases. It identifies 'if-then' rules that indicate which items tend to
appear together.

• Apriori Algorithm: Finds frequent itemsets in transaction data and generates association
rules based on support, confidence, and lift metrics.
• FP-Growth Algorithm: More efficient than Apriori; uses a tree-based structure to mine
frequent patterns without generating candidate itemsets.

4. Anomaly / Outlier Detection

Anomaly detection identifies data points that deviate significantly from expected patterns. Since
anomalies are rare and often unlabeled, unsupervised methods are ideal for detecting them.

How Unsupervised Learning Works — Example Walkthrough (K-Means Clustering)

1. Initialize K cluster centroids randomly in the feature space.

2. Assignment Step: Assign each data point to the nearest centroid based on Euclidean
distance.
3. Update Step: Recompute each centroid as the mean of all points assigned to its cluster.
4. Repeat Steps 2–3 until convergence (centroids no longer move significantly).
5. Output: K cluster assignments that reveal the natural groupings within the data.

Applications of Unsupervised Learning

Domain Application Technique Used

E-commerce & Retail Customer segmentation — K-Means / Hierarchical

grouping customers by Clustering
purchasing behavior for
targeted marketing campaigns

Page 17 of 19
Unit 3 — Machine Learning | Study Notes

Finance Fraud and anomaly detection DBSCAN / Isolation Forest

— identifying unusual
transaction patterns that may
indicate credit card fraud
Healthcare Gene expression clustering — Hierarchical Clustering / PCA
grouping patients with similar
genetic profiles to discover
disease subtypes
Natural Language Processing Topic modeling — discovering LDA (Latent Dirichlet Allocation)
latent themes in large
collections of documents (news,
research papers)
Retail / Supermarkets Market basket analysis — Apriori / FP-Growth
'Customers who buy bread and
butter also tend to buy jam'
(product placement,
promotions)
Computer Vision Image segmentation — K-Means / Mean Shift
partitioning pixels into
meaningful regions without
labeled training data
Cybersecurity Network intrusion detection — Anomaly Detection /
flagging unusual traffic patterns Autoencoders
that deviate from normal
behavior
Recommender Systems Collaborative filtering — Matrix Factorization / Clustering
grouping users with similar
taste profiles to recommend
content
Social Media Community detection — Graph Clustering
identifying hidden communities
or groups within social networks
Manufacturing Predictive maintenance — DBSCAN / Autoencoders
detecting anomalous sensor
readings indicating equipment
about to fail

Detailed Application Examples

Example 1 — Customer Segmentation in E-Commerce

An e-commerce company has millions of customer records including purchase history, frequency of
visits, average order value, product categories browsed, and geographic location. Using K-Means
clustering, the company groups customers into meaningful segments such as:

Page 18 of 19
Unit 3 — Machine Learning | Study Notes

• High-value loyal customers: frequent buyers with high spending — target with premium
loyalty programs
• Bargain hunters: purchase primarily during sales — target with discount alerts
• New customers: recently acquired — target with onboarding offers and tutorials
• At-risk customers: previously active but lapsing — target with win-back campaigns

Without any labeled data, the ML system reveals these business-critical groupings purely from
behavioral patterns in transaction data.

Example 2 — Anomaly Detection in Banking

A bank processes millions of credit card transactions daily. The vast majority are legitimate, but a
small fraction are fraudulent. Since fraudulent transactions are rare and take novel forms, they
cannot be comprehensively labeled in advance.
An unsupervised anomaly detection model (using Isolation Forest or Autoencoders) learns the
statistical distribution of normal transactions and flags transactions that deviate significantly as
potential fraud — enabling real-time intervention without requiring a fully labeled fraud dataset.

Example 3 — Topic Modeling in News Classification

A news aggregator receives thousands of articles daily from diverse sources. Without manual
categorization, Latent Dirichlet Allocation (LDA) — an unsupervised topic modeling algorithm —
automatically discovers latent themes such as 'Politics,' 'Sports,' 'Technology,' and 'Economy' by
analyzing word co-occurrence patterns across documents, enabling automatic content organization.

Advantages of Unsupervised Learning

• Does not require expensive, time-consuming manual data labeling
• Can discover hidden patterns and structures that humans may not have anticipated
• Applicable to any domain where raw, unlabeled data is available
• Excellent for exploratory data analysis and hypothesis generation

Limitations of Unsupervised Learning

• Results are harder to evaluate objectively since there are no ground-truth labels to compare
against
• Discovered clusters or patterns may not always be meaningful or actionable in a business
context
• Sensitive to feature scaling, distance metrics, and algorithm hyperparameters (e.g., choice
of K in K-Means)
• Generally less accurate than supervised learning for specific prediction tasks where labeled
data is available

Page 19 of 19