Learning with Answers
For three chapters Iris has been collecting pieces without a name for the whole. Features describe each case; a label holds the right answer; training shrinks the misses; the held-out test set keeps everyone honest. Then, in a planning meeting, a data scientist sums up an entire project in one word — "it's supervised, the comfortable kind" — and the whole room nods. Iris writes the word down and goes to find out what it means.
It means less than it sounds like it should. Supervised learning is machine learning from examples that carry the right answer — and that is the family everything in this book has belonged to so far. The delivery model learned from past orders that included the actual minutes; the fake-review idea assumes reviews marked genuine or fake. This page makes the word precise, traces where the answers come from, and shows why this one family does most of ML's paid work.
The Recipe, Restated
Line the pieces up and the definition writes itself. You have a table of past cases. Feature columns describe each one; a label column records what the right answer turned out to be. Training tunes the model until its outputs match those recorded answers, and the test set checks the skill on rows it never saw. Nothing here is new — what is new is seeing that the label column is doing a special job. It is the correction the model learns against. That correction is the "supervision"; no person watches anything.
Think of flashcards. A question on the front, the answer on the back; skill grows by guessing, flipping, and being corrected a few thousand times. Supervised learning is exactly the flashcard setup — the backs of the cards exist, written into the data before training starts. Hold that picture one chapter longer: Chapter 5 asks what could possibly be learned from cards with blank backs, and the answer is surprisingly rich. From here on, though, the term is supervised learning, and it simply means the answers were in the table.
Where the Answers Come From
Topic 08 sorted data sources into two piles, and those piles now reveal what they really were: the supervision supply chain. Some answers reality writes for free. Predict a delivery time, and the truth arrives forty minutes later — the courier taps "delivered", the app records the actual minutes, and a fresh labeled example drops into the table without anyone lifting a finger. Churn works the same way: wait a month and every customer has honestly answered "stayed" or "left" with their behavior.
Other answers have to be manufactured. No log anywhere says "this review was fake" — people must read reviews and mark them, one by one, thousands of times, and that labeling is real paid work. So the first budget question of any supervised project is not "which model?" but "where will the answers come from, and at what price?" A project whose labels arrive free scales effortlessly; a project that needs a human verdict for every training row carries that cost before a single parameter is tuned.
Why It Rules Business ML
Now the claim from the opening — that this family does most of ML's paid work — stops being mysterious. A business is a machine for generating recorded input-and-outcome pairs. Every order ends in an actual delivery time. Every transaction ends up charged back or clean. Every customer, given time, stays or leaves. Wherever history has been writing answer sheets, supervised learning applies directly — no new philosophy required, just the table and the loop.
This is also why the reflex question from Chapter 2 — "where would the labels come from?" — turns out to be the master question. Asking it is asking "is this a supervised problem, and can we afford the supervision?" When the labels exist, the path ahead is the well-worn one this chapter walks. When they don't, the idea either pays for labeling, waits for reality to write answers, or moves to Chapter 5's territory.
Its Two Faces
One distinction remains, and it organizes the next two pages. Look at the label column and ask what shape the answer is. Sometimes it is a kind: genuine or fake, fraud or clean, pizza or pasta. Sometimes it is an amount: 38 minutes, 214 orders, a $4 tip. Kind or amount — every supervised problem lands on one side of that line.
The two sides get their own names — classification for kinds, regression for amounts — and their own pages, because the shape of the answer changes what a miss looks like and how honesty is measured. But keep hold of what they share: the same table, the same training loop, the same held-out exam. Two faces, one family.
- "Supervised means a human supervises the training." Nobody watches. Supervision means only that the training examples carry the right answers — the label column does all the correcting.
- "The model is told the rules." It is told only the answers. The rules — which feature patterns lead to which answers — are exactly what it must work out for itself during training.
- "Supervised learning is one specific algorithm." It is a family defined by the with-answers setup. Many different models live inside it — Topic 19 tours three, and Topic 20 shows a committee of hundreds.
- "If we have data, we have supervision." Only if the answer column exists or can be created. A million unlabeled reviews supervise nothing — labels are the ingredient, and sometimes the whole cost.
- "Supervised" is in every job post, tool description, and data-team sentence from here on — it now means something exact to you: the answers were in the table.
- Your "where would the labels come from?" reflex has found its home: it is the question that decides whether an idea is a supervised problem, and often whether it is affordable at all.
- The kind-or-amount split you just learned is the first thing a data scientist does with any feature request — after the next two pages, you'll do it in seconds too.
Knowledge Check
What makes a machine-learning setup "supervised"?
- A person monitors the training process and corrects the model by hand
- Engineers write down the rules the model must follow
- The training examples include the right answers to learn against
- The model's results are checked on a test set after training
For delivery-time prediction, where do the training labels come from?
- Paid labelers read through each order and estimate how long it probably ended up taking
- Reality records them — each delivery's actual time is logged when it completes
- The current model's predictions are saved and reused as labels
- Customers fill in a survey about how long their order took
Plateful wants a model that answers "which cuisine will this new customer order first?" — and one that answers "how much will their first order cost?" How do the two problems compare?
- Both are supervised; one predicts a kind, the other a number
- Only the cost question is supervised — cuisine choices can't be labeled
- They are identical problems, since both use the same customer table
- Each needs a different family of ML, because the answers differ
Why does supervised learning dominate business ML?
- Supervised models are mathematically guaranteed to be more accurate than any alternative
- Supervision is free, since labels never cost anything to produce
- It is the only kind of machine learning that actually works
- Business history keeps recording input-and-outcome pairs — ready-made answer sheets
You got correct