Learning Objectives
By the end of this lesson, the learner should be able to:
- Define supervised learning and explain its fundamental characteristics.
- Define unsupervised learning and explain its fundamental characteristics.
- Distinguish between supervised and unsupervised learning approaches.
- Explain the role of features, labels and target variables in supervised learning.
- Identify common supervised learning tasks and algorithms.
- Identify common unsupervised learning tasks and algorithms.
- Explain classification and regression as supervised learning problems.
- Explain clustering as a major unsupervised learning technique.
- Determine when supervised or unsupervised learning is appropriate for a business problem.
- Evaluate the business benefits and limitations of each learning approach.
1. Introduction to Learning Approaches
In Lesson 8.1, machine learning was introduced as a method through which computational systems learn patterns from data.
However, not all machine learning problems are the same.
One of the most fundamental distinctions in machine learning is between:
- Supervised learning
- Unsupervised learning
The distinction primarily concerns whether the training data contains a known outcome or target that the model is expected to learn.
A simplified comparison is:
Supervised Learning
Data + Known Outcome → Learn Relationship → Predict Outcome
Unsupervised Learning
Data without Known Outcome → Discover Patterns or Structure
Understanding this distinction is important because the choice of learning approach depends heavily on the nature of the business problem and the available data.
2. Supervised Learning
2.1 Definition
Supervised learning is a machine learning approach in which an algorithm learns from historical data containing known outcomes or labels.
The model is given examples showing both:
- The input information
- The desired output
The objective is to learn a relationship that can be applied to new observations.
For example, a bank may have historical loan records containing:
- Customer income
- Credit score
- Existing debt
- Loan amount
- Employment status
- Repayment history
- Default status
The model can learn from these historical examples to predict whether a new loan applicant is likely to default.
The known historical outcome—default or non-default—is the target variable.
3. The Structure of Supervised Learning
A simplified supervised learning dataset can be represented as:
|
Customer |
Income |
Credit Score |
Loan Amount |
Default |
|
C001 |
60,000 |
720 |
300,000 |
No |
|
C002 |
35,000 |
590 |
250,000 |
Yes |
|
C003 |
80,000 |
760 |
400,000 |
No |
In this example:
- Income = feature
- Credit Score = feature
- Loan Amount = feature
- Default = target/label
The algorithm uses historical relationships between the features and the target to construct a model.
The model can then be applied to a new customer:
|
Customer |
Income |
Credit Score |
Loan Amount |
|
C004 |
50,000 |
650 |
280,000 |
The model might estimate:
Probability of default = 27%
The bank can then incorporate this prediction into its lending process.
4. Main Types of Supervised Learning
Supervised learning is commonly divided into two major categories:
1. Classification
Classification predicts a category or class.
Examples:
- Fraud / Not Fraud
- Churn / No Churn
- Approved / Rejected
- High Risk / Medium Risk / Low Risk
- Spam / Not Spam
2. Regression
Regression predicts a continuous numerical value.
Examples:
- Sales revenue
- Property price
- Customer lifetime value
- Monthly demand
- Loan amount
- Delivery time
Therefore:
Classification → Predict a category
Regression → Predict a numerical value
These approaches will be examined in greater detail in Lesson 8.3.
5. Examples of Supervised Learning in Business
Supervised learning is particularly useful when an organization has historical data containing known outcomes.
Banking
A bank could predict:
- Loan default
- Credit risk
- Fraudulent transactions
Telecommunications
A telecommunications company could predict:
- Customer churn
- Customer upgrade probability
- Service failure
Retail
A retailer could predict:
- Product demand
- Customer purchases
- Customer response to promotions
Insurance
An insurer could predict:
- Claim probability
- Claim cost
- Policy risk
Human Resources
An organization could analyze:
- Employee turnover
- Workforce requirements
- Recruitment outcomes
The key requirement is that there is generally some historical outcome against which the model can learn.
6. Advantages of Supervised Learning
Supervised learning offers several important advantages.
6.1 Clear Business Objective
Because the target variable is known, the business problem can often be defined precisely.
6.2 Measurable Performance
The model’s predictions can be compared with known outcomes.
For example, if a model predicts that 100 customers will churn, actual customer behavior can later be examined to assess performance.
6.3 Strong Decision Support
Supervised learning can support specific business decisions such as:
- Who is likely to leave?
- Which customers are likely to buy?
- Which transactions are potentially fraudulent?
- How much revenue is expected?
6.4 Automation
Once appropriately developed and deployed, models can generate predictions for large numbers of observations much faster than manual analysis.
7. Limitations of Supervised Learning
Supervised learning also has limitations.
7.1 Requires Historical Outcomes
A model needs appropriate historical examples containing known outcomes.
If a business has never recorded customer churn, for example, it may be difficult to build a conventional supervised churn model.
7.2 Label Quality
Incorrect or inconsistent labels can negatively affect model performance.
7.3 Historical Bias
Historical decisions may contain biases that become embedded in the training data.
7.4 Changing Conditions
A model trained using historical relationships may become less effective if market conditions change.
For example, customer behavior during an economic crisis may differ substantially from historical behavior.
8. Unsupervised Learning
8.1 Definition
Unsupervised learning is a machine learning approach in which the algorithm works with data without a predefined target or label.
Rather than being told what outcome to predict, the algorithm attempts to discover meaningful patterns, structures, groups or relationships within the data.
A simplified representation is:
Input Data → Algorithm → Discovered Structure
For example, a retailer may have customer information including:
- Age
- Spending
- Purchase frequency
- Product categories
- Average transaction value
However, the retailer may not know how many meaningful customer groups exist.
An unsupervised learning technique can analyze the data and identify groups of customers with similar characteristics.
9. Clustering
One of the most common forms of unsupervised learning is clustering.
Clustering involves grouping observations based on similarity.
For example, an organization might discover three customer groups:
Segment A: High-value frequent customers
- High spending
- Frequent purchases
- Long customer relationships
Segment B: Occasional customers
- Moderate spending
- Infrequent purchases
- Moderate engagement
Segment C: Low-engagement customers
- Low spending
- Infrequent purchases
- Limited interaction
These groups were not necessarily predefined by management.
The algorithm identified patterns within the customer data.
Clustering and segmentation will be explored in greater detail in Lesson 8.4.
10. Other Applications of Unsupervised Learning
Unsupervised learning can be used for:
Customer Segmentation
Discovering groups of customers with similar characteristics.
Anomaly Detection
Identifying observations that appear substantially different from normal patterns.
For example, an unusual transaction might be flagged for further investigation.
Pattern Discovery
Identifying relationships or structures that were not previously known.
Document and Text Analysis
Grouping documents based on similarity of content.
Product Analysis
Identifying products that tend to be associated with similar purchasing patterns.
11. Example: Customer Segmentation
Consider an online retailer with 100,000 customers.
The organization collects:
- Annual spending
- Number of purchases
- Average transaction value
- Website visits
- Product categories purchased
Management wants to understand the customer base but has no predefined customer categories.
An unsupervised learning algorithm could analyze these characteristics and identify several groups.
For example:
Cluster 1: High spending, frequent purchases
Cluster 2: Low spending, frequent purchases
Cluster 3: High spending, infrequent purchases
Cluster 4: Low spending, infrequent purchases
Marketing managers can then examine these groups and determine appropriate strategies.
The important point is that the algorithm discovers the structure, while business professionals interpret the meaning and decide how to act on it.
12. Advantages of Unsupervised Learning
12.1 Discovery of Unknown Patterns
It can reveal structures that management did not previously know existed.
12.2 Useful When Labels Are Unavailable
Organizations often possess large datasets without predefined target variables.
Unsupervised learning can still extract value from such data.
12.3 Exploratory Analysis
It is particularly useful during the early stages of understanding a dataset.
12.4 Customer Segmentation
Businesses can identify naturally occurring customer groups rather than imposing arbitrary categories.
13. Limitations of Unsupervised Learning
Unsupervised learning also presents challenges.
13.1 Interpretation
An algorithm may identify a group, but management still needs to determine what that group means from a business perspective.
13.2 No Single “Correct” Answer
Different algorithms or settings may produce different groupings.
13.3 Evaluation Can Be Difficult
Because there is no predefined target, evaluating whether a discovered structure is genuinely useful can be more complicated.
13.4 Patterns May Not Have Business Value
An algorithm can identify statistically interesting patterns that do not lead to useful business decisions.
Therefore, business interpretation remains essential.
14. Supervised versus Unsupervised Learning
The distinction can be summarized as follows:
|
Dimension |
Supervised Learning |
Unsupervised Learning |
|
Target variable |
Present |
Usually absent |
|
Main purpose |
Predict known outcome |
Discover structure/patterns |
|
Learning data |
Labeled |
Unlabeled |
|
Common tasks |
Classification, regression |
Clustering, pattern discovery |
|
Evaluation |
Often directly measurable against known outcomes |
Often more exploratory |
|
Example |
Predict customer churn |
Segment customers |
|
Business question |
“Who is likely to churn?” |
“What types of customers exist?” |
15. A Practical Business Comparison
Consider a telecommunications company.
Management has two different questions.
Question A
Which customers are likely to cancel their subscriptions within the next three months?
This is a supervised learning problem because historical churn outcomes are available.
The model can learn from customers who previously churned and those who remained.
Question B
What distinct types of customers exist within our customer base?
This may be an unsupervised learning problem because management may not have predefined customer categories.
The algorithm can identify groups based on customer characteristics and behavior.
Thus:
Predicting a known outcome → Supervised
Discovering unknown structure → Unsupervised
16. Semi-Supervised Learning
Between supervised and unsupervised learning is another approach known as semi-supervised learning.
Semi-supervised learning uses:
- A relatively small amount of labeled data
- A larger amount of unlabeled data
This can be useful when obtaining labels is expensive or time-consuming.
For example, a financial institution may have:
- 5,000 transactions manually classified as fraudulent or legitimate
- 5 million transactions without labels
Instead of ignoring the large unlabeled dataset, semi-supervised approaches can potentially use information from both datasets.
This approach can be particularly valuable in environments where expert labeling requires substantial time and resources.
17. Self-Supervised Learning
Another important concept in modern machine learning is self-supervised learning.
In self-supervised learning, the system generates learning signals from the data itself rather than relying entirely on manually assigned labels.
This approach has become particularly important in areas such as:
- Natural language processing
- Computer vision
- Speech processing
- Large-scale representation learning
For example, a language model can learn relationships within text by using parts of the available text to predict other parts.
Self-supervised learning is conceptually different from conventional supervised learning because the training signals are derived from the underlying data rather than being manually assigned business labels.
18. Choosing Between Supervised and Unsupervised Learning
The choice should begin with the business question.
Use supervised learning when:
- There is a clearly defined target.
- Historical outcomes are available.
- The organization wants to make predictions.
- Performance can be evaluated against known outcomes.
Examples:
Will this customer churn?
Is this transaction fraudulent?
What will next month’s sales be?
Use unsupervised learning when:
- There is no clearly defined target.
- The organization wants to discover patterns.
- Customer or product groups are unknown.
- Exploratory analysis is required.
Examples:
What customer segments exist?
Which products have similar behavior?
Are there unusual patterns in our transactions?
19. Business Analyst’s Role in Selecting the Approach
Selecting supervised or unsupervised learning should not be left solely to technical specialists.
The business analyst should help clarify:
Business objective
What decision or problem needs to be addressed?
Available data
What information exists?
Target variable
Is there a reliable historical outcome?
Decision requirement
Does management need a prediction or a discovery?
Business value
What action will be taken based on the result?
Risk
Could the model’s output affect customers, employees or other stakeholders in significant ways?
This reinforces a fundamental principle:
The analytical method should follow the business problem, not the other way around.
20. Combining Supervised and Unsupervised Learning
In real business environments, organizations do not always need to choose only one approach.
The approaches can complement one another.
For example, a bank could first use unsupervised learning to identify different customer segments.
It could then develop supervised learning models within those segments to predict:
- Product purchases
- Loan default
- Customer churn
- Credit behavior
This can create a broader analytical workflow:
Discover Customer Segments
↓
Understand Segment Characteristics
↓
Build Predictive Models
↓
Generate Predictions
↓
Design Business Actions
Such combinations can provide deeper insights than relying on a single analytical approach.
21. International Business Example: Amazon
Large digital businesses such as Amazon illustrate how machine learning can support different types of business problems.
A supervised learning approach can be used conceptually for problems where historical outcomes are available, such as estimating the likelihood of a customer purchasing a particular product.
Unsupervised techniques can support discovery-oriented tasks such as identifying patterns among customers, products or behavioral data.
The broader lesson is that large organizations often employ multiple machine learning approaches simultaneously, depending on the business question.
22. International Business Example: Banking and Fraud Detection
Financial institutions face millions of transactions and must identify potentially suspicious activity.
A supervised model could be trained using historical transactions that have been investigated and classified as:
- Fraudulent
- Legitimate
The model could then estimate the likelihood that a new transaction belongs to either category.
Unsupervised techniques can also be useful when fraud patterns are unknown. Anomaly-detection approaches may identify transactions that differ significantly from normal customer behavior.
This demonstrates an important distinction:
Known fraud patterns → Supervised learning
Previously unknown unusual patterns → Potentially unsupervised/anomaly-detection approaches
In practice, financial institutions may combine multiple methods.
23. Challenges in Real-World Business Data
Choosing a learning approach becomes more complicated when organizational data is imperfect.
Common problems include:
Missing labels
Historical outcomes may not have been recorded.
Incorrect labels
Records may have been classified incorrectly.
Imbalanced outcomes
A fraud dataset, for example, may contain millions of legitimate transactions but relatively few confirmed fraudulent transactions.
Changing patterns
Customer behavior may change over time.
Data silos
Relevant information may be distributed across different systems.
Privacy restrictions
Some variables may not legally or ethically be appropriate for modeling.
Therefore, choosing between supervised and unsupervised learning is only one part of the analytical process.
24. Practical Decision Framework
A business analyst can use the following framework:
Step 1: Define the business question
What decision needs to be improved?
Step 2: Identify the desired output
Is the objective to predict an outcome or discover patterns?
Step 3: Examine the data
Is a reliable target variable available?
Step 4: Select the learning approach
- Known target → Supervised learning
- No predefined target → Unsupervised learning
- Limited labels + substantial unlabeled data → Consider semi-supervised approaches
Step 5: Evaluate business usefulness
Will the result lead to a meaningful business action?
Step 6: Monitor results
Does the model or discovered structure remain useful as business conditions change?
Reflection Exercise
A retail bank wants to improve customer analytics.
It has the following information:
- Customer age
- Account balance
- Transaction frequency
- Loan history
- Product ownership
- Digital banking activity
- Customer complaints
- Historical loan default status
Management has three questions:
- Which customers are likely to default on a new loan?
- What different types of customers exist in the bank’s customer base?
- Which customers are likely to purchase a new investment product?
Reflection Questions
For each question:
- Would you recommend supervised or unsupervised learning?
- What would be the target variable, if applicable?
- What features could be used?
- What business decision could be supported?
- What potential risks should management consider?
Best Practices
Organizations applying supervised and unsupervised learning should follow these practices:
- Begin with the business question.
- Determine whether a reliable target variable exists.
- Assess the quality of labels before using supervised learning.
- Do not assume that every discovered cluster represents a meaningful business segment.
- Combine technical analysis with business interpretation.
- Use appropriate evaluation techniques for the selected approach.
- Consider bias and representativeness in historical data.
- Avoid using sensitive or inappropriate variables without proper justification and governance.
- Validate analytical findings with domain experts.
- Translate model outputs into actionable business decisions.
- Monitor performance as business conditions change.
- Document why a particular learning approach was selected.
Lesson Summary
Supervised and unsupervised learning represent two fundamental approaches to machine learning.
Supervised learning uses labeled historical data containing known outcomes. It is primarily used to predict outcomes through tasks such as:
- Classification
- Regression
Examples include predicting:
- Customer churn
- Loan default
- Fraud
- Sales
- Demand
Unsupervised learning works primarily with data without predefined target labels. It is used to discover:
- Groups
- Patterns
- Relationships
- Anomalies
- Structures
A major application is clustering and customer segmentation.
The key distinction is:
Supervised learning asks: “What outcome can we predict?”
Unsupervised learning asks: “What patterns or structures can we discover?”
The lesson also introduced semi-supervised and self-supervised learning, which extend the range of approaches available for modern analytical problems.
Ultimately, selecting the appropriate learning approach should depend on the business objective, data availability, target definition, decision requirements and expected business value.
References
- IBM — Supervised vs. Unsupervised Learning
IBM: Supervised vs Unsupervised Learning - Google for Developers — Machine Learning Glossary
Google Machine Learning Glossary - Microsoft Learn — Machine Learning Concepts
Microsoft Learn: Machine Learning - scikit-learn — Supervised Learning
scikit-learn: Supervised Learning - scikit-learn — Unsupervised Learning
scikit-learn: Unsupervised Learning - OECD — Artificial Intelligence
OECD AI Policy Observatory