Skip to content

P1-8.1 Supervised Learning: Input and Label

Section ID: P1-8.1 Version: v2026.07.23

Chapter 7 dealt with search spaces, computational limits, and heuristics. Now we move into learning types. The first baseline is supervised learning.

The core of this section is not an algorithm, but the label. To understand supervised learning, we first need to ask what the label points to. If we treat a label vaguely as the answer, we can easily miss the more important issue: what standard people attached to the data in the first place.

In Korean explanation, it is often safer to understand a label as something closer to an identification tag than as an absolute statement of truth. A label is a marker attached so that the data can be distinguished. Labeling is the act of attaching such markers. Explanations of data labeling from AWS and IBM also describe it as identifying raw data and assigning labels to it. That marker tells the model what standard it should follow, but it does not automatically mean the label is absolute truth in every sense.

So the central question of this section is:

what does a label point to,
and how does the model try to follow that label standard?

Supervised learning uses examples in which input and label are given together. The model is trained so that when a new input arrives, it produces an output that follows the label standard. So in 8.1, the main object of attention is not the complex inside of the model, but the meaning and standard of the labels attached to the data.

supervised learning is not the memorization of a list of labels; it is the matching of relationships between inputs and labels

In Part 1, this section fixes the basic distinction among supervised learning, label, labeling, target, classification, and regression. The basic distinction between input and output was handled first in 4.2, and the execution flow of training and prediction was handled in 5.1 and 5.2. Since Chapter 7 already organized search and heuristics, this section can focus directly on the structure created by labeled data.

supervised learning, label, labeling, target, classification, and regression can initially sound like similar data-processing terms. A short positional split looks like this:

Term Very short meaning Role in this section
supervised learning learning from examples that contain labels the first baseline in Chapter 8
label a target marker attached to data the standard the model tries to fit
labeling the act of attaching labels to data the preparation that builds the learning standard
target the value the model is meant to match often close in meaning to label
classification predicting a categorical label one representative form of supervised learning
regression predicting a numeric label another representative form of supervised learning

The minimum distinction to keep is this: supervised learning is learning with labels, a label is a target marker, labeling builds the standard, classification chooses a name tag, and regression predicts a number.

The basic distinction between input and output was already fixed in 4.2, and the difference between training and inference was handled in Chapter 5. Here we add one layer on top: how examples with labels become a training signal.

The working definition here is:

supervised learning is the method of training a model to predict outputs for new inputs by using examples where input and label are given together

Setting a Learning Standard with Inputs and Labels

  • Understand what a label indicates in supervised learning.
  • Read a label as something closer to an identifiable marker than to an absolute answer.
  • Explain labeling as the act of attaching distinguishing markers to data.
  • Understand supervised learning as using examples that contain both input and label.
  • Distinguish the relation among label, target, and expected output.
  • Distinguish classification from regression at an introductory level.
  • Understand that the presence of labels does not always mean absolute truth.
  • Prepare to keep supervised learning, unsupervised learning, and reinforcement learning separate.

Three Standards

This section is not mainly about algorithms. It is about understanding what a label is. If the following three standards are fixed, the whole flow becomes much clearer.

Standard Why it matters Level of understanding needed here
supervised learning uses examples that contain both inputs and labels This is the most basic standard that separates it from other learning types. It is enough to understand that people-attached distinguishing markers come together with past examples.
a label is more accurate when read as an identification marker than as absolute truth This reduces the mistake of treating labels as unquestionable answers. Keep the intuition that labels are markers attached to distinguish cases.
the model tries to generalize that label standard to new inputs This summarizes the goal of supervised learning in one sentence. Understand it as an attempt to apply standards seen in past examples to new ones.

A Label Is a Signal That Tells the Model What Direction to Learn

A label is the value attached to data that tells the model what output it is expected to match. In supervised learning, the word supervised does not mean a person gives direct instructions at every step. It is closer to meaning that labels inside the training data serve as signals that guide learning.

For example, if customer-support sentences are labeled as delivery, refund, or exchange, the model is trained to match the relationship between the sentence and that label. If price data includes actual sale prices, the model is trained to match the relationship between input conditions and that numeric label.

So a label is not just a name tag. More precisely, it is a distinguishing marker attached within a problem definition. In that sense it resembles a symbol in the broad sense that people assign a sign such as delivery, refund, or exchange to distinguish one type of case from another. But labeling is not the same thing as symbolic AI. Labeling is a preparation step for machine learning.

a label is a distinguishing marker attached to data
labeling is the act of attaching that marker
the marker is then used as the target output the model is trained to match
the label standard determines the direction of learning

Supervised Learning Starts from Labeled Examples

Google’s Machine Learning Glossary explains supervised learning as training a model from input features and the labels associated with them. It also explains that a labeled example contains one or more features and a label, and that such examples are used for training.

At this stage, we do not need to go deeply into features. The first important structure is simpler:

input + label = labeled example

For customer-support classification:

Input Label
I want a refund. refund
When will the delivery arrive? delivery
The item arrived broken. exchange
Can I cancel the payment? refund

The model sees these examples and then predicts which label fits a new sentence:

new input: The product I ordered still has not arrived.
model output: delivery

The crucial point is that supervised learning uses a standard already attached by people. Without labels, the same problem would have to be approached through unsupervised learning or another method. So when we read a supervised-learning problem, we should ask about the label standard before we ask about the model type.

Label, Target, and Expected Output

In supervised learning, the label is the target output the model is meant to match. The scikit-learn glossary explains target as the dependent variable passed as y to an estimator's fit method in supervised and semi-supervised learning, and also lists label as another name.

The practical reading used in this book is:

Expression Meaning in this section
label the distinguishing marker attached to data and used as target output
labeling the act of attaching labels to data
target the value the model is meant to match; often nearly the same as label
expected output the output that people expect the model to produce

In introductory explanation, labels are often called answers. But this wording should be read carefully. A label is usually a distinguishing marker attached within a specific problem definition and data-collection standard.

Consider the sentence:

I want to cancel the payment.

Its label may differ depending on the labeling scheme:

Labeling scheme Possible label
workflow handling cancellation
customer intent refund
payment-system classification payment

So a label is not an absolute truth floating outside context. It is a marker attached inside a specific problem definition and used as a target value for supervised learning.

a label is a distinguishing marker attached inside a problem definition
in supervised learning, that marker is used as the target output
if the problem definition changes, the label for the same input may also change

The Basic Flow of Supervised Learning

The basic flow can be summarized like this:

collect labeled examples
let the model predict outputs from inputs
compare predictions with labels
train the model so the gap becomes smaller
use the trained model to predict for new inputs

Using customer-support examples:

Stage Explanation
data preparation collect past support sentences and human-assigned labels
training adjust the model so it matches the relation between sentences and labels
prediction when a new sentence arrives, output a label
evaluation check whether the prediction fits the actual business standard

As Chapter 5 explained, training is the procedure that adjusts internal model values. In 8.1, the point is to see why this labeled-data setting is called supervised learning.

Classification Predicts Categorical Labels

Classification is a supervised-learning problem where the input is assigned to one or more predefined categories or labels. The key is that the label is categorical. Google’s glossary distinguishes classification models, which predict classes, from regression models, which predict numeric values.

Customer-support type selection is a classification problem:

Input Possible label
The delivery is late. delivery
I want a refund. refund
The product is damaged. exchange

Typical classification outputs are category labels:

Problem Input Label
spam classification email title and body spam, normal
support-message classification support sentence delivery, refund, exchange, other
image classification image cat, dog, car
risk detection transaction information normal, suspicious

Regression Predicts Numeric Labels

Regression is a supervised-learning problem where the model predicts a continuous numeric value. Here the label is not a category name but a number. Google’s glossary also explains that regression models predict numeric values.

Examples:

Problem Input Label or target
delivery-time prediction region, distance, order time, carrier state expected duration
product-price prediction product attributes, season, inventory, demand price
server-cost prediction request count, data transfer, instance count monthly cost
temperature prediction time, region, pressure, humidity temperature

Classification and regression are both supervised learning. The difference is the form of the label:

Distinction Label form Example
classification category delivery, refund, exchange
regression number 3.5 hours, 12,000 won, 27.2 degrees

The introductory summary is:

classification chooses a name tag
regression predicts a number

In Supervised Learning, Label Standards Matter

In supervised learning, labels are the learning signal. So if the label standard is unstable, the model also becomes unstable.

Labeling is the act of attaching distinguishing markers to data. That can look like simple data entry, but in reality it is the act of deciding what counts as the same thing and what should be distinguished.

If different people attach different labels to the same support sentence, the result becomes unstable:

Input Label A Label B Problem
I want to cancel the payment. refund cancellation intent standard and workflow standard are mixed
I want to change the address. delivery address change the label system overlaps
The item arrived broken. exchange defective item handling method and cause classification are mixed

Such data still satisfies having labels, but it is not necessarily good supervised-learning data, because the standard the model is meant to fit is inconsistent.

Questions to ask when designing labels:

does the label match the work objective?
can different people apply the same standard?
do labels overlap with one another?
can the model really predict the label from the actual available input?

Whenever possible, the label standard should be documented together with the dataset. The Datasheets for Datasets proposal is helpful here because it argues that motivation, composition, collection process, and recommended use should all be documented so that the dataset is understandable to later users. At the level of 8.1, the key point is simply that label standards are part of dataset context.

Having a Label Does Not Automatically Mean Having Truth

Supervised learning is often explained as learning from data with the correct answer. As an introductory metaphor, that can be useful, but it should not be read too literally. The more careful phrase is learning from data that has labels.

Some labels are relatively clear:

Input Label
image of a handwritten digit one of 0 through 9
product sales record actual sale price
delivery-completion record complete or incomplete

But other labels depend on human judgment, policy, historical context, or measurement devices:

Input Why the label may be unstable
customer emotion classification people may interpret emotion differently
risky-transaction judgment it can change with policy and cost criteria
harmful-content classification it can change with culture, law, and platform policy
quality evaluation raters may apply different standards

Google’s glossary also notes that human raters can make mistakes during labeling, and that proxy labels must be chosen carefully when the real label is not directly available.

So the important attitude in supervised learning is this:

the existence of a label does not mean truth is automatically settled
first check what standard created the label

The Boundary Between Supervised Learning and Other Learning Types

Section 8.1 covers only supervised learning. It helps to fix the boundary to the next sections in advance.

Learning type Data and signal Central question Relation to the next sections
supervised learning inputs and labels what label should be predicted for this input? 8.1
unsupervised learning unlabeled data what structure exists inside the data? 8.2
reinforcement learning states, actions, and rewards what action increases long-term reward? 8.3

Deep learning does not live on the same conceptual level as this table. Deep learning is an approach that learns representations with multi-layer neural networks, and it can be used together with supervised learning, unsupervised learning, or reinforcement learning.

Checklist

  • I can explain supervised learning as a learning method that fits the relation between inputs and outputs from examples that contain both input and label.
  • I can distinguish classification and regression as predicting categorical labels and numeric labels.
  • I can explain how label, target, and expected output are used at the same level.
  • I can explain that the label standard should match the business goal and should be attachable consistently by different people.
  • I can explain that the presence of labels does not mean absolute truth is present.
  • I can distinguish supervised learning, unsupervised learning, and reinforcement learning by input-label pairs, structure discovery, and reward signals.
  • I can explain that the label standard and data context should be checked before the model.

Sources and Further Reading