Chapter Five · No Label Column

Learning Without Answers

Everything so far learned from examples with the answers attached. This chapter removes the answer column and asks what a computer can still discover: the natural groups hiding in a customer table, the one order that doesn't fit the pattern of thousands — and, in a short honest visit, the third family of learning, where a machine teaches itself by trial and reward.

4 topics

Every model in Chapter 4 was trained the same way: rows of examples, each with the answer attached. Churned or stayed. Fake or genuine. Delivered in 38 minutes. The label column was the teacher — the model's whole job was to predict it. But some of the most valuable questions at Plateful come with no label column at all. "What kinds of customers do we have?" has no answer column, because nobody knows the kinds in advance. That is exactly why marketing is asking.

This chapter is about what data can teach when nobody attached the answers. First the idea itself — unsupervised learning, and the honest catch that comes with having no answer key. Then its two workhorses: clustering, which finds the natural groups in a table, and anomaly detection, which spots the point that doesn't fit — the 3 a.m. order of forty desserts on a fresh card. The chapter closes with the third and strangest family, reinforcement learning, where the machine generates its own examples by acting. And the fraud detector built here walks straight into Chapter 6, which opens by asking the uncomfortable question: how do you score it?

The same customer table — with and without the answer column

With a label column

Orders per month, usual hours, average spend — and a final column: "churned? yes/no". A model can learn to predict that column. This was Chapter 4.

No label column at all

The same rows, and the answer column simply isn't there. Nothing to predict — so what can be learned anyway? Groups, odd ones out, hidden patterns. This chapter.

Topics in This Chapter

Topic 21

Learning Without Labels

Marketing asks what kinds of customers Plateful even has — and the table has no answer column. Unsupervised learning: discovering groups, oddities, and patterns nobody defined, and the honest catch of having no answer key.

Concept

Topic 22

Clustering

Five customer segments nobody invented — the machine found them in the data. How similarity becomes piles, why humans do the naming, and the how-many-groups dial that usefulness, not correctness, decides.

Concept

Topic 23

Finding the Odd One Out

Forty desserts at 3 a.m., new address, fresh card. How a model learns the shape of normal from ordinary behavior alone — and why a flag is a raised eyebrow, never a verdict.

Concept

Topic 24

Learning by Trial and Reward

The third family: nobody hands the learner examples at all — it acts, gets rewarded or stung, and improves. Why it beat Go champions, why your data team probably doesn't use it, and the human-feedback trick that returns in Chapter 9.

Concept