The Three Types of Machine Learning

If "machine learning" sounds like one big thing, that's because that's how it's marketed. In practice, almost every ML system in the world falls into one of three buckets: supervised, unsupervised, or reinforcement learning. Knowing which bucket a problem belongs to is half the battle — it tells you what data you need, what tools to use, and what answers you can expect.

This lesson walks you through all three with real examples you'll recognize from your own apps.

What You'll Learn

• The three families of machine learning, in plain English
• How to spot which type a problem needs
• Real-world examples for each (Spotify, Netflix, ChatGPT, self-driving cars)
• A simple test you can run with ChatGPT to classify any ML problem

Type 1: Supervised Learning (Learning from Examples with Answers)

Supervised learning is the most common kind. The "supervisor" is a labeled dataset — examples that come with the correct answer attached.

Examples:

• Spam detection — emails labeled "spam" or "not spam"
• House price prediction — historical sales data with the actual sale price
• Medical diagnosis — X-rays labeled "tumor present / not present" by doctors
• Customer churn prediction — past customers labeled "stayed" or "left"

The model studies thousands of labeled examples, finds patterns linking inputs to outputs, and uses those patterns to predict new cases. Two flavors:

• Classification — predicting a category (spam / not spam, dog / cat, fraud / legitimate)
• Regression — predicting a number (price, temperature, sales)

The tell: if you can imagine listing examples with a "right answer" column, it's supervised learning.

Type 2: Unsupervised Learning (Finding Hidden Patterns)

Sometimes you have data but no labels. You don't know the "right answer" — you just want to see what's in there. That's unsupervised learning.

Examples:

• Customer segmentation — group shoppers into types without knowing the types in advance
• Spotify "Daily Mixes" — group your liked songs into vibes
• Anomaly detection — find unusual patterns in network traffic or transactions
• Topic discovery — group news articles by theme

The two main jobs:

• Clustering — group similar items together
• Dimensionality reduction — squish complex data into something simpler to visualize

The tell: if you're saying "I have data, what's interesting in it?" — that's unsupervised.

Type 3: Reinforcement Learning (Learning by Trial and Error)

Reinforcement learning is the closest thing to how a child learns to ride a bike. The system tries actions, gets rewards or penalties, and gradually learns a strategy.

Examples:

• AlphaGo — Google's Go-playing AI, learned by playing millions of games against itself
• Self-driving cars — learn from rewards (smooth driving) and penalties (collisions in simulation)
• Recommendation systems — Netflix's homepage shuffles based on what you actually click
• Robot vacuum routing — learns the most efficient path through your apartment
• Trading algorithms — buy / sell actions rewarded with profit

There's no labeled dataset. There's a goal (maximize a reward) and a feedback loop.

The tell: if the system needs to make a sequence of decisions and the right action depends on the situation, reinforcement learning fits.

What About ChatGPT and Claude?

Modern large language models like ChatGPT, Claude, and Gemini use all three types stitched together:

1. Unsupervised pre-training on massive text data (find patterns in language)
2. Supervised fine-tuning on examples of good answers
3. Reinforcement learning from human feedback (RLHF) to make answers more helpful

This is why they feel so capable — they're not one ML technique, they're a stack of them.

A Practical Sorting Test

Open ChatGPT, Claude, or Gemini. Try this prompt:

"I'll describe three real-world problems. For each one, tell me whether it's a supervised, unsupervised, or reinforcement learning problem, and explain why in one sentence:

1. Predicting whether a credit card transaction is fraudulent
2. Grouping our customer support tickets into common themes we haven't named yet
3. Training a drone to fly through an obstacle course"

You should get: 1) supervised classification, 2) unsupervised clustering, 3) reinforcement learning. Now invent three problems from your own life or studies and test yourself.

How This Course Will Use Each Type

Type	Where you'll see it	Tool we'll use
Supervised	Predicting house prices, classifying images	Google Sheets, Teachable Machine
Unsupervised	Grouping survey responses	ChatGPT / Claude
Reinforcement	Discussion only (most no-code tools don't expose RL)	Conceptual examples

You don't need to memorize this table. Just know that 90% of business ML problems are supervised — that's where you'll spend most of your career using ML.

Common Beginner Mistake

People often try to apply supervised learning when they don't have labels yet. If you only have raw data and no "right answer" column, you're either doing:

• Unsupervised learning (find structure)
• Or you need to create labels first (often manually)

Don't skip the labeling step — bad labels create worse-than-useless supervised models.

Key Takeaways

• Supervised learning needs labeled examples and predicts known categories or numbers
• Unsupervised learning finds patterns in unlabeled data, often by clustering similar items
• Reinforcement learning trains agents through trial-and-error rewards
• ChatGPT and Claude combine all three under the hood
• Knowing the type tells you what data you need and what answers you can expect

Next up: data — the actual fuel that makes any of this possible. Garbage data in means garbage predictions out, no matter how fancy the model.