1. Supervised learning

Supervised learning means training a model using historical examples where the correct answer—the label—is already known.

Example dataset:

Here:

Features = Premium, ClaimCount, Complaints
Label = Renewed

The model learns a relationship:

Historical features + known labels
                    ↓
                 Training
                    ↓
                  Model
                    ↓
New features without label → Predicted label

It is called “supervised” because the correct historical answers guide the learning process.

This does not mean a human watches every training step. It means labeled examples provide supervision.

2. Classification

Classification is a supervised-learning task where the model predicts a category.

Binary classification

Binary means two possible classes.

Examples:

Renew / Not renew
Fraud / Not fraud
Pass / Fail
Spam / Not spam
High risk / Normal risk

Possible ML.NET label:

public bool Renewed { get; set; }

A binary-classification model can return:

PredictedLabel: false
Probability:    0.81
Score:          model-specific numeric value

Multiclass classification

Multiclass classification predicts one category from more than two possible categories.

Examples:

Support department:
Billing / Technical / Sales / General

Document type:
Invoice / Contract / Identity Proof / Certificate

Candidate issue:
Connectivity / Authentication / Proctoring / Content

Possible C# label:

public string Category { get; set; } = string.Empty;

Important distinction:

Two categories      → Binary classification
Three or more       → Multiclass classification

3. Regression

Regression is supervised learning where the model predicts a numeric value.

Examples:

Expected premium amount: ₹18,750
Estimated resolution time: 42 minutes
Expected milk production: 1,250 litres
Property price: ₹75,00,000

Possible C# label:

public float ResolutionMinutes { get; set; }

Classification versus regression:

A number used as an identifier is not a regression target.

CustomerId = 1058

Although numeric, this is an identity—not a quantity to predict.

4. Unsupervised learning

Unsupervised learning works with data that does not contain a known label.

Example:

There is no column saying:

CustomerSegment = ?

Instead, the algorithm searches for useful patterns or groups in the features.

5. Clustering

Clustering is an unsupervised-learning task that groups similar records.

A clustering model might discover:

Cluster 1:
Low premium, no claims, high renewal history

Cluster 2:
High premium, few complaints, long relationship

Cluster 3:
Frequent claims, frequent complaints, low renewal history

The model generally returns identifiers such as:

ClusterId = 1
ClusterId = 2
ClusterId = 3

These numbers do not automatically have business meaning.

A domain expert must examine the clusters before assigning names such as:

Stable customers
High-value customers
Service-risk customers

Never assume:

Cluster 1 = best
Cluster 3 = worst

Cluster IDs are merely group identifiers.

6. Other unsupervised tasks

Unsupervised approaches can also help with:

• Finding unusual records
• Reducing many variables into fewer dimensions
• Discovering relationships
• Exploring data before labeling
• Identifying similar documents or customers

Later, we will separately study:

• Anomaly detection
• Embeddings
• Similarity search
• Vector databases

7. Reinforcement learning

Reinforcement learning, or RL, trains an agent through interaction with an environment.

New terms:

• Agent: The decision-making system.
• Environment: The system or world the agent interacts with.
• State: The current situation visible to the agent.
• Action: A choice the agent can make.
• Reward: Numeric feedback indicating how useful an action was.
• Policy: The strategy the agent learns for selecting actions.

Example:

State:
Candidate is struggling with algebra questions

Possible actions:
Give easier question
Give same-level question
Give harder question
Show a hint

Reward:
Positive if learning improves
Negative if candidate abandons or performance declines

The agent repeatedly experiences:

Observe state → Choose action → Receive reward → Update policy

8. RL is different from supervised learning

In supervised learning:

Input → Correct answer is already available

In reinforcement learning:

State → Agent chooses action → Consequence produces reward

The system may not know the single correct action in advance. It learns which actions produce better accumulated rewards.

9. Immediate versus long-term reward

Suppose an adaptive learning system always gives easy questions.

The student answers correctly and receives an immediate positive result. But the student may learn very little.

A good reward may need to consider:

• Knowledge improvement
• Completion
• Engagement
• Difficulty balance
• Long-term retention

This is why reward design is difficult.

A badly designed reward can teach unintended behaviour.

10. When reinforcement learning is appropriate

RL may be considered when:

• Decisions happen sequentially.
• One action affects future situations.
• A reliable reward can be measured.
• Exploration can occur safely.
• A simulation or controlled environment exists.
• Long-term outcomes matter.

RL is usually inappropriate when:

• One prediction is required.
• Historical labels already exist.
• Experimental actions could harm users.
• Reward is vague or delayed beyond measurement.
• Deterministic business rules solve the problem.
• A safe training environment is unavailable.

For most early business ML projects, supervised learning is more common than reinforcement learning.