Data, Features, and Labels: The Three Words Behind Every Model

Why this matters

Every model you have used, from a spam filter to a song recommender to a chatbot, was built from the same three things: data, features, and labels. Get them wrong and the model fails quietly, and nobody can tell you why.

The big idea on one slide

Data is the pile of examples. Features are the clues. Labels are the answers.

  • Data: the rows of real examples you collected.
  • Features: the details you wrote down about each one.
  • Labels: the correct answer you attached to each one.

How it actually works

Data just means examples kept in a list, like the camera roll on your phone. Each photo is one example. On its own, a pile of photos is useless, because you have not decided what to ask about.

Features are the clues you choose to write down. Think of a doctor’s checklist: temperature, cough, sore throat. The doctor ignores your haircut and your shoe size. Features are the columns you keep, and the wrong columns teach the model the wrong lesson.

Labels are the right answers, stuck to the examples. It is a worksheet with the answer key stapled to the back. For every photo of a dog, a person wrote dog. Without labels, a model can sort things into groups, but it cannot tell you what those groups mean.

Here is the surprising part: features and labels matter more than the math. A simple method with good features beats a fancy method with bad ones. That is like packing for a trip. A perfect suitcase does not help if you packed for the beach and you are going skiing.

Slide-by-slide talking points

Slide 1: The hook

  • Imagine a model that decides whether you get a loan, and nobody can explain why it said no.
  • That happens when the features and labels were chosen carelessly, or never chosen on purpose at all.
  • Tonight we open the box and name the three parts inside every model.

Slide 2: The big idea

  • Data is the pile of examples, features are the clues, and labels are the answers.
  • Remove any one of the three and you cannot train a model at all.
  • Remember that sentence and you can follow almost any AI conversation.

Slide 3: How it works

  • First you collect examples, then you decide which details to write down as features.
  • Next you attach the correct answer to each example, and those answers are the labels.
  • The model hunts for patterns linking features to labels, then guesses on new examples it has never seen.

Slide 4: A worked example

  • Let us predict whether a club member shows up to Saturday practice.
  • Our features are whether they replied in the group chat and how many days away practice is.
  • After five examples, the model guesses that anyone who replied yes to a practice two days away will show up.

Slide 5: The takeaway

  • When a model does something dumb, check the features and labels before blaming the math.
  • Labels are human choices, so they carry human opinions and human blind spots.
  • Whoever picks the labels shapes the model more than whoever writes the code.

Worked example

Here is a made-up example, small enough to check by hand. Your club wants to guess whether a member will show up to Saturday practice.

You collect five past Saturdays. Two features per person: did they reply yes in the group chat, and is practice two days away or five days away? The label is what actually happened.

  • Replied yes, two days away: showed up.
  • Replied yes, two days away: showed up.
  • Replied yes, five days away: did not show up.
  • Replied no, two days away: did not show up.
  • Replied no, five days away: did not show up.

Step one: find the rows that match the new person. A new member replied yes, and practice is two days away. Two rows match, and both are showed up. Step two: predict from the matching rows, so the guess is shows up.

Step three: question the result. Five rows is tiny, and replying yes might just be politeness. Real models use thousands of rows, but the logic is the same. Better features and cleaner labels give better guesses.

Try it this week

Spend ten minutes with Teachable Machine, a free browser tool from Google. Choose the image project and train a model to tell two objects apart, say your phone and your hand. Hold each object up to the webcam and click train to add it as a class. Your classes are the labels, and the pixels the camera sees are the features.

Then break it on purpose. Cover half the object, or move to a dim room, and watch the confidence scores wobble. You just found a feature the model leaned on too hard.

Question for the room

If a model learns from labels, who decides what the right answer is?

The point to reach: labels are choices made by people, and two people would label the same comment differently. Questions like whether a comment counts as toxic are arguments about values, not about code. So the person labeling the data shapes the model as much as the engineer does.

Words to know

  • Data – examples collected in a list, one row per example.
  • Feature – a detail you wrote down about an example, like a column in a spreadsheet.
  • Label – the correct answer attached to an example.
  • Training set – the labeled examples you use to build the model.
  • Model – the pattern-finder that turns features into a guess.

Go deeper