The Three Families

Every algorithm in this course belongs to a family. Know the family, and you instantly know what data it needs, and what it can do. There are three. And the difference comes down to one simple question, does the data come with answers?

• Every algorithm has a family
• It decides what data you need
• One question separates them

The three families

The first family is supervised learning. The data comes with answers, called labels, like a house and its true sale price. The second is unsupervised learning. There are no answers. The algorithm finds hidden structure by itself, like groups of similar customers. The third is reinforcement learning. Nobody gives answers. Instead, an agent tries actions, and earns rewards, like a program learning to win a game. Most business value today comes from the first family, so that is where we spend most of our time.

Inside supervised learning

Supervised learning splits into two kinds of problem. In regression, the answer is a number, like a price, a temperature, or next month's sales. In classification, the answer is a category, like spam or not spam, or which digit is in an image. A simple test is to ask, could the answer be halfway between two others? A price of two hundred and fifty thousand makes sense. But an email cannot be half spam, at least not as a final answer.

Name the family

Let's practise. Predicting a house price is supervised regression. Deciding if a tumour is cancerous is supervised classification. Grouping shoppers by behaviour, with no labels, is unsupervised clustering. And teaching a robot to walk is reinforcement learning. By the end of this course, you will build the first three yourself.

Summary