Skip to content

P1-8.2 Unsupervised Learning: Structure and Representation

Section ID: P1-8.2 Version: v2026.07.20

Section 8.1 explained a label as a distinguishing marker attached to data. Supervised learning tries to match the relationship between inputs and outputs using examples where inputs and labels are given together.

This section starts from the opposite side. Unsupervised learning is a learning method that tries to find structure, clusters, or representations inside data without labels attached in advance by people.

The central question is:

when there are no labels,
what is the model trying to find inside the data?

unsupervised learning is not learning without any standard at all; it is learning to find structure inside data without human-provided labels

This section organizes unsupervised learning, structure, clustering, dimensionality reduction, and cluster labels under one viewpoint: learning from unlabeled data by finding structure. The baseline sense of label and supervised learning was fixed in 8.1, and the baseline sense of representation was introduced earlier in 3.3 and 4.3.

This section does not calculate specific unsupervised-learning algorithms. K-means, hierarchical clustering, PCA, t-SNE, and autoencoders pass by only as names and roles.

This section also first closes what the model is trying to find inside data when there are no labels. Distance functions, the curse of dimensionality, and cluster-evaluation metrics return in the clustering and dimensionality-reduction chapters of Part 4, and the detailed mathematics of manifold learning is not introduced here.

This section also does not redefine label at length. Its basic distinction was already fixed in 8.1. Here the focus is on what becomes the learning standard when there are no labels. Representation is also not a new term introduced for the first time here. We only connect that unsupervised learning can change representations to reveal structure.

The working definition here is:

unsupervised learning is a method that finds structure, clusters, or representations inside data without labels attached by people

Finding Structure and Representation Without Labels

  • Explain unsupervised learning as learning that finds structure in unlabeled data.
  • Understand that no labels does not mean no goal or no standard.
  • Distinguish clustering from dimensionality reduction at an introductory level.
  • Avoid confusing cluster labels with human-attached labels.
  • Avoid mixing unsupervised learning with supervised learning, reinforcement learning, or deep learning as if they were one category.

Three Standards

Standard Why it matters Level of understanding needed here
unsupervised learning tries to find structure inside data without labels This is the most basic distinction from supervised learning. It is enough to understand that the model looks for similarity, grouping, or hidden axes without human answer tags.
the result does not come with automatic meaning; people still have to interpret it This reduces the mistake of reading clusters or reduced dimensions as truth by themselves. Even if the model makes groups, people still have to decide what those groups mean.
clustering, dimensionality reduction, and similar tasks ask different questions This prevents unsupervised learning from being treated as one single technique. It is enough to distinguish what is similar?, how can we reduce dimensions?, and where is structure dense?

A short role separation is useful:

Term Very short meaning Role in this section
unsupervised learning learning without labels by finding structure the second baseline of Chapter 8
structure relationships and patterns inside data the central concept of what is being sought
clustering grouping similar things together one representative unsupervised-learning problem
dimensionality reduction looking at many dimensions through fewer ones another representative problem that makes structure easier to see
representation the form in which data appears inside the model the angle through which structure becomes visible
cluster label a group number or marker produced by the algorithm something that must be kept separate from human-provided supervised labels

No Labels Does Not Mean No Goal

When people hear unsupervised, they sometimes imagine that no guidance exists at all. The more accurate point is narrower: there are no labels attached in advance by people.

there are no human-provided labels in advance
but there is still a goal: find structure in the data

Google’s Machine Learning Glossary explains unsupervised learning as training a model to find patterns in unlabeled examples. If supervised learning fits outputs to given labels, unsupervised learning looks at the arrangement and relationships of the data itself without such labels.

Suppose we have customer-support sentences like these:

Support sentence Human-provided label
I want a refund. none
When will the delivery arrive? none
The item arrived broken. none
Can I cancel the order? none
I want to change the address. none

No labels such as delivery, refund, or exchange are attached. But similarities may still exist among the sentences. Unsupervised learning tries to find that kind of similarity, repeated pattern, or hidden axis.

Structure Means Relationships Inside Data

In unsupervised learning, structure means that data is not just scattered randomly. It has some relationship or pattern inside it.

Examples:

Data Structure that might be found
customer purchase history groups of customers with similar buying tendencies
product browsing records products frequently viewed together
support sentences groups of sentences with similar intent
server logs request patterns that differ from normal behavior
image features groups of visually similar shapes or textures

The key point is that nobody labeled these in advance as group A or delivery inquiry. The algorithm tries to find structure through distances, similarities, densities, or variances in the data.

supervised learning fits to labels
unsupervised learning looks for structure without labels

Clustering Groups Similar Things Together

Clustering is one of the most representative examples of unsupervised learning. scikit-learn explains that unlabeled data can be clustered through sklearn.cluster.

Clustering in the AI sense is not just a person manually sorting items by intuition. A person can, of course, create groups manually in statistics or business work. But in machine learning, clustering is closer to a procedure that represents data numerically through features and then groups it using calculable criteria such as distance, similarity, density, or variance.

Distinction Human manual grouping Algorithm-based clustering
basis human reading, business experience, prior knowledge chosen features, distance, similarity, density, variance
result a human-assigned business group or category name cluster number, centroid, or grouping structure
strength easy to reflect context and intention directly easier to inspect large data with a consistent calculable standard
caution standards may differ between people the grouping may not match business meaning directly

So clustering is not the machine automatically understands the true categories. It is closer to the system groups things that look similar under the chosen representation and criteria.

For example, K-means divides data into a chosen number of clusters and repeatedly assigns each point to the nearest centroid. Stanford CS229 notes describe k-means as a problem of creating cohesive clusters from unlabeled data, and explain that each training example is assigned to the nearest cluster centroid. scikit-learn also explains that k-means divides samples into several groups and tries to minimize the within-cluster sum-of-squares criterion.

Input data What clustering tries to find
customer purchase history groups of customers with similar patterns
news articles groups of articles with similar themes
sentence embeddings groups of sentences with similar meaning
image features groups of visually similar images
sensor records groups of similar operating states

The important question is always:

by what notion of similarity are we saying these belong together?

The answer can change with the chosen features. Even with the same data, the grouping may change depending on what we compare.

Situation Features we might compare Possible clusters Caution
shopping-mall customers purchase frequency, average order amount, revisit interval frequent buyers, occasional high spenders, window shoppers names such as VIP or churn risk are later human interpretations
support messages sentence embeddings, frequent terms, message length delivery-like messages, refund-like messages, defect-like messages clusters may not match business departments exactly
news articles title and body embeddings, named people, places, topic terms political articles, economic articles, sports articles articles about the same event may mix with articles about the same topic
product images color, shape, texture, feature vectors products with similar color tone, form, or shooting angle the result may differ from the category people expect
server logs request time, status code, latency, route path normal request patterns, slow request patterns, error-heavy patterns these are anomaly candidates, not the root-cause analysis itself

Using support-message embeddings as an example:

Cluster Possible included sentences
cluster 1 When will delivery happen?, My order still has not arrived.
cluster 2 I want a refund., Can I cancel the payment?
cluster 3 The item arrived broken., Can I exchange it?

The algorithm did not start with the labels delivery, refund, or exchange. It grouped the sentences first by similarity. A human may then look at the group and say, this cluster seems close to delivery inquiries.

That is why a cluster label such as cluster 1 is not the same thing as the label discussed in supervised learning.

a cluster label is a grouping marker created by the algorithm
it is not the same as a human-attached supervised-learning label

Dimensionality Reduction Views Complex Data Through Fewer Axes

Unsupervised learning is not only about grouping. It is also used to look at complex data through fewer axes. This is called dimensionality reduction.

Here, a dimension can be thought of as one descriptive axis of the data. An image may have many pixel-related dimensions, a document may have many token or embedding dimensions, and customer data may have many behavioral variables.

Dimensionality reduction tries to view that high-dimensional data through fewer dimensions while preserving important structure as much as possible.

Data What dimensionality reduction can help with
high-dimensional image features seeing whether similar images lie near each other
sentence embeddings observing groups of semantically similar sentences
customer behavior data finding major behavioral axes for visualization
sensor data inspecting normal patterns versus unusual ones

scikit-learn’s explanation of PCA describes it as decomposing multivariate data into orthogonal components that explain variance. At the introductory level, the key point is simpler:

dimensionality reduction means viewing complex data again through fewer axes
while trying to preserve important structure

Representation Is the Form in Which Data Appears to the Model

The word representation also matters here. Representation is the form into which original data is transformed so that the model or algorithm can handle it more easily.

For example, sentences are difficult to compute over as raw language. When turned into numeric vectors, distances or similarities between sentences can be calculated. Images are arrays of pixels, but inside models they may appear through features related to shape, color, boundary, or texture.

In unsupervised learning, the model may be used not to fit labels, but to find or build representations that make data structure visible.

Original data Example representation
support sentences sentence embeddings
product images image feature vectors
customer behavior logs behavioral-pattern vectors
document collections lower-dimensional topic-related representations

Unsupervised learning can also connect to making structure visible by changing the representation without labels. Later chapters connect that more directly to deep learning and embeddings.

Unsupervised-Learning Results Need Interpretation

Because no labels are attached in advance, it is risky to read unsupervised-learning results directly as business truth.

Suppose a clustering result looks like this:

Cluster Included customers
cluster 1 high purchase amount, high visit frequency
cluster 2 many visits but low purchasing
cluster 3 purchases only in a particular season

It would be too quick to declare these immediately as VIP, churn-risk, or seasonal customer. Those names are interpretation added later by people. Unsupervised learning can reveal candidate structure, but the business meaning of that structure still has to be examined.

Useful questions include:

by what criteria was this structure created?
would a different distance function or variable choice change it?
is the result interpretable in business terms?
is the human name attached to it an over-interpretation?

The Boundary from Other Learning Types

Chapter 8 separates three basic learning types. Summarized again:

Learning type Data and signal Central question Place in this chapter
supervised learning inputs and labels what label should be predicted from this input? 8.1
unsupervised learning unlabeled data what structure exists inside the data? 8.2
reinforcement learning states, actions, rewards which actions increase reward over time? 8.3

Deep learning is not on the same axis as this table. It can be used inside unsupervised learning, but it is not identical to unsupervised learning itself.

Checklist

  • Explain unsupervised learning as a way of finding structure in unlabeled data.
  • Explain why no labels does not mean no goal.
  • Explain clustering as a way of grouping similar things together.
  • Explain dimensionality reduction as a way of looking again at complex data through fewer axes.
  • Distinguish a cluster label from a human-attached supervised label.
  • Explain why unsupervised-learning results still require human interpretation and review.
  • Explain that what counts as similar, what structure we are trying to find, and who interprets the result should be separated.
  • Explain why cluster numbers or reduced coordinates should not be read directly as business meaning or truth.

Sources and Further Reading