Learn AI /How Machines Learn /Unsupervised learning
Unsupervised learning
- Intermediate
- How Machines Learn
- Technical readers
Back to the milestone you came from
In five seconds
What it is
A way of training a model on unlabelled data to identify possible structure, similarities, groups or unusual cases without target answers for each example.
In plain words
The system looks through many examples without answer tags and tries to find what tends to go together, what is different and what structure may be hidden in the data.
An everyday example
At a logistics depot, workers receive hundreds of parcels from Nairobi, Lima, Bangkok and Tunis with no shelf labels telling them where each one belongs. They begin by comparing size, weight, shape and handling marks, then place similar parcels on nearby shelves. After a while, clear groups appear: small urgent envelopes, heavy machine parts, fragile glass items and a few parcels that do not fit anywhere well. The workers did not have an answer sheet in advance, but they still uncovered structure in the pile.
Real unsupervised learning does not literally inspect objects the way people do, and it does not inherently know what the groups mean. A model may find mathematical patterns that are not useful in practice, or it may miss the categories a human cares about. Also, some unsupervised methods do not create clear groups at all; they may compress data or estimate density instead.
Remember thisUnsupervised learning does not get answer keys; it searches unlabelled data for structure that people can then interpret and use.
Understand it
Understand it
Why it matters
Many real datasets arrive without labels because labelling is costly, slow or impossible at scale. Unsupervised learning helps people find patterns before they know exactly what to ask for. That can support early warning, data exploration, monitoring, search, segmentation and later model design. For decision-makers, the key point is that it can reveal structure in large datasets, but the discovered groups or anomalies still need human interpretation and validation. A pattern found by the model is not automatically meaningful, fair or useful for the task at hand.
How unlabelled data becomes structure
After viewing this, a reader should be able to say that unsupervised learning starts with examples without target answers and produces discovered structure such as groups, embeddings or anomaly scores.
01
1. Input data
Many training examples enter the system with features but no labels or ground truth answers.
Draw a panel of 12 to 16 dots or cards, each with tiny feature bars or numbers, and no category tags. Add the caption: 'Unlabelled examples'.
02
2. Compare patterns
The algorithm measures similarity, distance or statistical regularity across examples.
Draw arrows from several input dots into a central box labelled 'compare features / find regularities'. Include small distance lines between some dots.
03
3. Adjust representation
During training, internal settings change so similar cases are placed closer, common structure is summarised, or rare cases stand out.
Inside the central box, show a simplified transformation: scattered dots entering, more organised dots exiting. Use one short note: 'internal parameters updated during training'.
04
4. Produce outputs
The result may be clusters, compact embeddings, or anomaly scores rather than named predictions.
On the right, split into three mini-panels: coloured groups of dots labelled 'clusters', a 2D map labelled 'embedding', and one isolated dot with a warning outline labelled 'anomaly score'.
A left-to-right flow diagram. Start with many unlabelled examples shown as dots with feature values. In the middle, an algorithm compares examples and adjusts its internal representation. On the right, the same data appears as grouped clusters, a compressed map, and one highlighted unusual point. Read from input to learned structure to practical output.
Hovering over each output mini-panel highlights which kinds of tasks it supports: exploration, compression or anomaly detection.
Describe this without the diagram
The diagram is a left-to-right process with four stages. First, it shows many unlabelled examples, each described by features but with no correct answer attached. Second, an algorithm compares the examples to find similarity or statistical regularity. Third, training changes the model’s internal parameters so the data becomes more organised in its learned representation. Fourth, the outputs appear in three forms: clusters of similar examples, compact embeddings that summarise examples with fewer numbers, and anomaly scores that mark unusual cases. The main message is that unsupervised learning begins without labels and ends with discovered structure, not with predefined target answers.
With reduced motion: A static four-panel flow chart shows unlabelled examples, a central pattern-finding stage, an adjusted internal representation, and three possible outputs: clusters, embeddings and anomaly scores.
In this approach, the training data usually has inputs but no labels. A label is the target answer attached to an example, such as “fraud” or “not fraud”. Instead of learning to match inputs to known answers, the model looks for patterns in the data itself.
What goes in can be many examples described by features. A feature is a measurable property such as temperature, word frequency or sensor vibration level. During training, the algorithm adjusts its internal settings to place similar examples near each other, compress data into simpler representations, or flag examples that do not fit the usual pattern.
What comes out depends on the method. It might produce clusters, which are groups of similar examples; embeddings, which represent examples as numeric vectors; or anomaly scores, which estimate how unusual an example looks compared with the training data. These outputs help people explore data, monitor systems or prepare later models.
This is often confused with clustering, but clustering is only one type of unsupervised learning. It is also different from supervised learning, where each training example has a known target, and from self-supervised learning, where labels are created automatically from the data for a specific prediction task.
How “Sorting mixed parcels in a storeroom” maps onto the real thing
| In the story | In the AI system |
|---|---|
| parcels with no shelf labels | training examples without labels |
| size, weight, shape and handling marks | features describing each example |
| placing similar parcels on nearby shelves | grouping similar data points or learning a representation |
| parcels that do not fit any shelf well | anomalies or outliers |
A worked example
A water utility monitors vibration and temperature from 500 pump readings each hour and wants to spot unusual behaviour before failure.
Starting from: Each reading has four features: vibration RMS, bearing temperature, power draw and flow rate. Example readings include A=(2.1, 61, 14.8, 120), B=(2.3, 60, 15.1, 118), C=(7.9, 83, 19.4, 92). No reading is labelled 'normal' or 'faulty'.
- Engineers train an unsupervised anomaly-detection model on two weeks of mostly routine readings.
- The model learns the usual combinations of feature values and assigns each new reading an anomaly score.
- Most new readings receive low scores because they resemble past operating patterns.
- Reading C receives a much higher score because its vibration and temperature are far from the common pattern while flow rate is lower than usual.
Result: If reading C receives a high score, the system flags it for inspection. The score does not prove a fault, but it gives staff a reason to check the pump early.
In a United Nations workplace
During a records retention review at a specialised agency, an ombudsman case officer must prepare a note for the file by close of business on an archive of case records linked to a technical division. The records have retention metadata but no pre-assigned topic or anomaly labels.
Within an approved secure environment, an unsupervised method could group records with similar permitted metadata and flag unusual patterns for review. It may reveal possible duplicates, recurring record types or outliers without being given target answers. The groups are analytical leads, not retention decisions.
Before you try thisCase records are highly confidential. The officer must not submit narratives, names, contact details, allegations or identifying metadata to any tool not expressly approved for ombudsman records. Even in an approved environment, access controls, purpose limitation and retention rules apply. Clusters or anomaly scores must not be treated as findings about a person or case.
Where you may hear it
- At an ICT governance board reviewing a proposal for anomaly detection
- In a vendor demo to a procurement unit, when the slide turns to clustering
- During first reading of a concept note for organising an unlabelled document archive
- At an evaluation office meeting on exploratory analysis of monitoring data
Avoid confusion
Avoid confusion
Often confused with
Supervised learning
A method for training a model on examples paired with target outputs, so it can predict outputs for new inputs.
Clustering
A method that groups similar examples together so patterns in data can be found without pre-written categories.
Self-supervised learning
A training approach that creates its own learning targets from raw data, so a model can learn useful patterns without manual labels.
Semi-supervised learning
A learning approach that combines labelled and unlabelled examples to train a predictor for the same target task.
Common misconceptions
Common misconception: It means the system learns with no human involvement.
Correction: People still choose the data, features, objective, algorithm and evaluation method. Human judgement is also needed to interpret whether discovered patterns are useful or misleading.
Common misconception: It always produces clear, correct categories.
Correction: Some methods do not create categories at all, and even clustering can produce groupings that depend strongly on feature choice, scaling and algorithm settings. The groups are mathematical patterns, not guaranteed real-world classes.
Common misconception: If there are no labels, there is nothing to evaluate.
Correction: Evaluation is harder, but not impossible. People can inspect stability, usefulness for downstream tasks, reconstruction quality, anomaly follow-up rates or agreement with external evidence collected later.
Quick check
Which situation best matches unsupervised learning?
Unsupervised learning works with unlabelled examples and looks for structure such as similarity groups or anomalies. Option 1 is supervised learning, option 3 is evaluation, and option 4 is reinforcement learning.
Go deeper
Go deeper
The simple version
The system looks through many examples without answer tags and tries to find what tends to go together, what is different and what structure may be hidden in the data.
The technical version
Unsupervised learning trains on samples x without paired target values y. The objective is not to minimise prediction error against known labels, but to model structure in the input distribution or in a learned representation of it. Depending on the method, this can mean estimating density, learning latent variables, reducing dimensionality, building embeddings, discovering clusters, or assigning anomaly scores. Common mechanisms include similarity or distance calculations, reconstruction objectives, probabilistic assumptions and representation learning. Outputs are often intermediate structures rather than final task labels, so their value is judged by interpretability, stability, compression quality, downstream utility or the practical usefulness of flagged outliers.
Reviewed external resources
EncyclopaediaUnsupervised learning (opens in a new tab)Wikipedia
This link gives a broad overview of the term, common methods and related concepts, and suits readers who want a general reference point.
Intermediatelink checked 2026-07-31
CourseMachine Learning Crash Course (opens in a new tab)Google
This course link suits readers who want a structured introduction to machine learning ideas and where unsupervised methods fit in practice.
Intermediatelink checked 2026-07-31
VideoStatQuest (opens in a new tab)StatQuest with Josh Starmer
This video source suits readers who learn well from step-by-step spoken explanations and worked intuition about core machine learning concepts.
Starterlink checked 2026-07-31
VideoVideo explanations: Unsupervised learning (opens in a new tab)YouTube search
This search results page helps readers browse multiple video explanations with different teaching styles, and suits someone comparing explanations rather than relying on one source.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Unsupervised learning (opens in a new tab)Reddit search
This search results page helps readers see common questions, confusions and practitioner discussions, and suits readers who want informal discussion alongside formal explanations.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
Unsupervised learning does not get answer keys; it searches unlabelled data for structure that people can then interpret and use.