Where Data Comes From

Two million rows. Iris does some skeptical arithmetic — at one row a minute, that's years of typing — and asks who entered it all. The answer: nobody. The app wrote almost every row itself. Each order already produces timestamps, an address, a restaurant, a courier's GPS trace, and recording them is simply the product doing its job.

That is the first and largest source of training data: records that pile up as a side effect of software working. But records alone don't make the table from Topic 06 — that table needed its answer column, and answers are not always free. This page follows the data back to its sources, and the labels back to theirs, because the two have very different price tags.

Products Record Themselves

Software keeps logs — running records of what happened, written automatically as the program works: this order was placed at 19:02, this courier's phone reported these coordinates, this payment was for this amount. Nobody creates logs for machine learning; they exist for receipts, debugging, and dashboards. The data team's craft begins with noticing that yesterday's operational records can answer today's prediction question.

A shop's security camera makes the same move. It was installed to catch shoplifters — yet the footage can also tell you when queues form, which shelves people linger at, how weather changes foot traffic. Nothing about the camera changed; someone asked the recordings a new question. Most of what gets called "big data" is exactly this: records collected for one purpose, reread for another.

Labels Reality Hands You

For delivery-time prediction, the label problem solves itself. The courier eventually arrives, the customer taps "received", and the true answer — 41 minutes — lands in the log about 40 minutes after the question was asked. Every completed order labels itself. The best labels in the business are the ones the world writes for you.

Plateful has more of these than it first appears. Did the customer order again within a month? Wait a month and look. Was the payment reversed as fraudulent? The bank's chargeback arrives on its own. Free labels share a profile: abundant, cheap, and delivered on reality's schedule, not yours — you can't hurry a month of waiting.

Labels Humans Must Create

Now take fake-review detection, the task from Chapter 1 with no writable rule. No log anywhere says "this review is fake." Reality never records that answer, because the answer is a judgment. Someone has to read the review, weigh it, and mark it — and then do that again several thousand times, or there is no answer column and no training.

That work is called labeling (or annotation): people creating the answers that reality didn't record. It is real, paid, skilled work — often the single most expensive line in an ML project's budget — and the labelers' judgment quality becomes the model's ceiling, since their marks are what it learns to imitate. Chapter 9 returns to labeling at a scale that may surprise you: armies of it sit behind the chatbots everyone uses.

Public and Bought Data

Not everything has to come from your own product. There are open datasets — collections published for anyone to use, from weather histories to millions of labeled photos assembled by researchers. There is data for sale, licensed from companies whose product produces it. And there are partnerships: Plateful could trade with a mapping company — traffic data in, delivery statistics out.

Every outside source drags in a quiet question: were we allowed to use this — by law, by the license, by the people the data describes? Chapter 10 gives that question the space it deserves. For now, carry the practical summary of this page: when someone proposes an ML feature, the fate of the whole project usually hides inside one innocent-sounding question — where would the labels come from?