Learning without answers

Data does not always contain correct answers. A store keeps the purchases of a million customers, but nobody has marked which customers are alike. A robot vacuum learns to drive around furniture, and the data holds no correct move for each situation either. These cases have two other problem settings.

Unsupervised

Unsupervised learning gets only the features, with no target. The model looks for structure in the data itself. Most often that structure is groups of similar samples—clusters.

A store can find, for example, a group of customers who buy only at night and only on sale. Another unsupervised task is finding unusual samples. A bank sends a payment unlike any of a million past ones to a manual check.

The model gives the groups no names. There is nothing to check them against either: the data has no target values. A person who looks at the result decides what each group means and whether it is useful.

Reinforcement

Reinforcement learning works like a game. A program takes an action in an environment and gets a reward—a number that says whether things got better or worse. The vacuum gets a plus for each cleaned meter and a minus for hitting a wall. The program's goal is to collect as much reward as possible over the whole run, not in one step.

There is no correct answer at each step, only the reward. It often comes with a delay: in chess, the result is known only at the end of the game. The program works out on its own which moves led to it. There is no data in advance: it appears from the program's actions and rewards.

Game bots and robots learn this way, and so does the choice of ad order on a website.

Common mistake

"Unsupervised" is read as "without a person", and the model is expected to set itself up. It does not. A person still chooses the features, the algorithm and its settings, such as the number of clusters. Unsupervised means one thing only: the data has no target.