Types of Machine Learning

The Three Families of ML

Not all machine learning is the same. Depending on what kind of data you have and what goal you’re trying to achieve, you’ll use different types of ML.

Think of it like teaching someone:

• Supervised = teaching with a tutor who gives feedback
• Unsupervised = learning by exploring on your own
• Reinforcement = learning by trial and error (like a video game)

1. Supervised Learning 🎓

In supervised learning, the training data comes with correct answers (called labels).

The model learns: “Given this input, what should my output be?”

Example: Predicting House Prices

The model learns the relationship between features (bedrooms, size, location) and the label (price). Then it can predict prices for new houses it hasn’t seen.

Two Subtypes

Classification — predict a category

Input: email text → Output: "spam" or "not spam"
Input: photo → Output: "cat", "dog", or "bird"
Input: patient data → Output: "has disease" or "doesn't have disease"

Regression — predict a number

Input: house features → Output: $450,000
Input: weather data → Output: 23.5°C temperature tomorrow
Input: ad data → Output: 3.2% click-through rate

Real-world supervised learning uses:

• Email spam filters
• Medical diagnosis assistance
• Credit card fraud detection
• Stock price prediction

2. Unsupervised Learning 🔍

In unsupervised learning, there are no labels — the model must find hidden patterns in the data on its own.

The model learns: “What natural groups or structures exist in this data?”

Example: Customer Segmentation

Imagine you run an online store. You have data about your customers:

Customer A: buys often, small amounts, mostly electronics
Customer B: buys rarely, large amounts, mostly luxury goods
Customer C: buys often, small amounts, mostly fashion

You don’t label them — you just let the ML model group similar customers together. It might discover:

• Group 1: Frequent small buyers (budget shoppers)
• Group 2: Rare big buyers (luxury shoppers)
• Group 3: Trend followers

Common Unsupervised Techniques

Real-world uses:

• Customer segmentation
• Recommendation systems
• Fraud detection (finding unusual transactions)
• Topic modeling in documents
• Gene expression analysis

3. Reinforcement Learning 🎮

In reinforcement learning, an agent learns by interacting with an environment and receiving rewards or penalties.

No labeled data. The agent just tries things and learns what works.

Agent takes action → Environment gives reward or penalty → Agent learns to maximize reward

The Classic Example: Learning to Play Chess

• Agent: The chess-playing AI
• Environment: The chess board
• Actions: Moving pieces
• Reward: +1 for winning, -1 for losing, 0 for each move
• Goal: Learn the strategy that wins most games

The agent starts playing randomly. Over millions of games, it discovers what moves lead to wins.

Real-world uses:

• Game-playing AIs (AlphaGo, OpenAI Five)
• Robot locomotion
• Autonomous vehicle control
• Trading bots
• RLHF in ChatGPT (teaching LLMs to be helpful)

Quick Comparison

Which Should You Use?

Do you have labeled data?
  ├── Yes → Supervised Learning
  │     ├── Predicting a category? → Classification
  │     └── Predicting a number? → Regression
  └── No → Unsupervised Learning
         └── Learning via interaction? → Reinforcement Learning