The Shape of an ML Problem

Almost every machine learning problem has the same basic shape: an input (the information you start with), an output (the thing you're actually trying to predict or produce), and training examples — real, past input-output pairs the system learns the pattern from.

Take a spam filter: the input is the text of an email, the output is a label — "spam" or "not spam" — and the training examples are thousands of real emails that were already correctly labeled by someone in the past. The system never gets a rulebook for what makes an email spammy; it works that out from the pattern across all those labeled examples.

This same input/output/examples shape holds up across wildly different problems: predicting a house's price (input: its features like size and location; output: a price; examples: past sale records), or recommending a video (input: what someone's watched before; output: a ranked list; examples: what people with similar history watched next). Once you can name these three parts, you can recognize an ML problem anywhere.

In a spam filter, what plays the role of "training examples"?

The thing a machine learning system is actually trying to predict or produce is called the ____.

A machine learning system is typically given an explicit rulebook defining exactly what makes an email spam.