Learn AI /How Machines Learn /Semi-supervised learning

Semi-supervised learning

Detail

In five seconds

What it is

A learning approach that combines labelled and unlabelled examples to train a predictor for the same target task.

In plain words

You have a few examples with correct answers and many without answers. This approach uses both, so the many unanswered examples can still help the model learn patterns.

An everyday example

A team in Geneva is organising a large box of printed paragraph pairs for a writing workshop. On only 20 pairs, a trainer has written labels such as “same meaning” or “different meaning”. Hundreds of other pairs have no labels, but by reading them, the team can still notice clusters: some use formal language, some are short summaries, and some differ only by a few words. They use the 20 labelled pairs to anchor the task, then use the larger pile to see which unlabelled pairs look similar to the anchored examples. This helps them sort new pairs more accurately than if they had used only the 20 labelled sheets.

Where the story stops working

People can consciously reason about the paragraphs and explain their choices. A machine-learning model usually does not reason in that human way; it adjusts numerical parameters from data. Also, real semi-supervised methods do not simply copy labels from the nearest-looking item. Different methods use different mathematical assumptions, and if those assumptions are wrong, extra unlabelled data can fail to help or can even hurt.

Remember thisSemi-supervised learning trains one predictor from both trusted labelled examples and additional unlabelled inputs. It can help when labels are scarce, but only when its assumptions and pseudo-labels hold up under evaluation.

Understand it

Why it matters

Many real projects have very little labelled data because expert annotation costs time and money. Semi-supervised learning can make better use of what an organisation already has: a small set of trusted examples and a much larger set of raw data. For decision-makers, the key question is not whether unlabelled data exists, but whether it is similar enough and clean enough to improve performance without reinforcing early mistakes.

How semi-supervised learning uses two kinds of data

After viewing, the reader should be able to explain how labelled and unlabelled examples are combined during training and why the result can outperform training on labelled data alone.

01
1. Small labelled set
Start with a limited number of examples that include the correct output label.
Draw a small stack of cards at top left with tiny tag icons. Label it 'Labelled examples: few but trusted'.
02
2. Large unlabelled set
Add many more examples without labels. They still contain useful patterns such as clusters, neighbourhoods or shared structure.
Draw a larger stack of cards at bottom left without tags. Label it 'Unlabelled examples: many but unanswered'.
03
3. Joint training
Feed both sets into one training stage. The method uses the labelled set for supervision and the unlabelled set to shape the model using similarity or confidence-based signals.
Place a central box titled 'Training'. Show two arrows entering it. Inside, include two short captions: 'fit known labels' and 'use structure in unlabelled data'.
04
4. Trained model
The output is a model that can assign labels or values to unseen inputs, often with better accuracy than using the labelled set alone.
Draw a model box on the right. Connect the training box to it with a bold arrow. Add a small note: 'better generalisation if assumptions hold'.
05
5. New input and prediction
A new unlabelled example goes through the trained model and receives a predicted output.
Add one card labelled 'new example' feeding into the model, with an arrow to a final card labelled 'predicted label/value'.

A left-to-right flow diagram with two inputs feeding one training process. The top input is a small labelled dataset. The bottom input is a large unlabelled dataset. Both enter a training stage that learns from labels and from data structure, then produce a trained model used to predict labels for new examples.

Hovering or tapping each stage highlights the specific role of labelled data, unlabelled data, and the final prediction path.

Describe this without the diagram

The diagram reads from left to right. At the top left is a small set of labelled examples, shown as a few cards with tags, to indicate trusted answers exist but are limited in number. At the bottom left is a much larger set of unlabelled examples, shown as many cards without tags, to indicate plenty of data but no answers. Both sets flow into a central training box. Inside that box, one note says the model learns from known labels, and another says it also uses structure in the unlabelled data, such as similarity or grouping. From this box, an arrow leads to a trained model. A new example then enters the model, and the model outputs a predicted label or value. The main message is that semi-supervised learning combines a small amount of labelled data with a large amount of unlabelled data during training.

With reduced motion: A static left-to-right diagram with two input stacks, one labelled and one unlabelled, both pointing to a training box, then to a trained model, then to a prediction output for a new example.

Many machine-learning systems need labelled examples. A label is the correct answer attached to an input, such as “sports” for a news article. In practice, labels are often expensive or slow to produce, while unlabelled data is plentiful.

Semi-supervised learning uses both kinds of data together. The model may learn from both datasets jointly, or use them in alternating steps. The aim is to use the extra structure in the unlabelled data to build a better predictor than labelled data alone would allow.

What goes in: a small labelled dataset and a larger unlabelled dataset. What changes inside: the model adjusts its internal parameters so similar inputs tend to get similar outputs, or so confident provisional labels on unlabelled items can help further training. What comes out: a model for tasks such as classification or regression.

It is often confused with self-supervised learning. They are related but not the same. In self-supervised learning, labels are created automatically from the data itself by a designed prediction task. In semi-supervised learning, at least some human-provided or otherwise external labels are already present and guide the learning.

How “Sorting paired paragraphs by meaning” maps onto the real thing

In the story	In the AI system
20 paragraph pairs with trainer-written tags	small labelled dataset
hundreds of unlabelled paragraph pairs	large unlabelled dataset
grouping by similarity in wording and structure	learning structure from unlabelled data
using tagged examples to guide later sorting	using labels to steer prediction on new inputs

A worked example

A reporting officer in a regional bureau is tagging public recommendations for inclusion in a progress report. A reviewed subset has topic labels, while the rest of the archive is unlabelled.

Starting from: The labelled and unlabelled records cover the same topic-classification task and have been separated from a held-out, human-labelled evaluation set.

Train an initial classifier on the human-labelled recommendations.
Apply it to the unlabelled archive and retain only high-confidence predictions as provisional pseudo-labels.
Retrain the classifier using the original human labels and the selected pseudo-labels, giving greater weight to the human-reviewed examples.
Test the revised classifier on the held-out evaluation set, including separate checks by language and document source.
Review pseudo-labelled clusters for systematic errors before using the tags to organise the progress report.

Result: If the unlabelled records form useful clusters around the reviewed categories, the revised classifier may outperform the labelled-only version. If its early predictions are biased or confidently wrong, pseudo-labelling can amplify those errors, so the reporting officer keeps the model only if the held-out evaluation supports it.

In a United Nations workplace

At a duty station with three working languages, a gender adviser is preparing the concept note for a side event in the margins of the session. To identify recurring themes in public event descriptions from previous sessions, the adviser uses a classifier trained on a reviewed subset tagged for participation, accessibility, financing and implementation. Semi-supervised learning also draws on the larger unlabelled archive to improve the same tagging task.

This is useful when colleagues can review only part of a large multilingual archive. The labelled records anchor the categories, while the unlabelled records may reveal clusters and language patterns that help the predictor separate those categories. The resulting tags still require human checking before they shape the concept note or talking points.

Before you try thisUse only public records in a tool approved for that material. Do not upload participant details, unpublished submissions or personal data to a service not approved for them. Check performance separately across all three working languages: pseudo-labels can reinforce early errors, especially where one language has fewer reviewed examples.

Where you may hear it

A vendor demo for multilingual document triage
An inter-agency working group reviewing an annotation plan
A procurement meeting discussing the volume of human-labelled examples
A first reading of a concept note for an automated classification pilot

Avoid confusion

Often confused with

Common misconceptions

Common misconception: Any large amount of unlabelled data will improve a model.
Correction: Not necessarily. Unlabelled data helps only if it is relevant to the task and the method's assumptions fit the data. If the unlabelled set comes from a different population, or the model makes biased early guesses, performance can stay flat or get worse.
Common misconception: Semi-supervised learning is just supervised learning with missing labels.
Correction: It uses missing labels in a deliberate way. In ordinary supervised learning, unlabelled examples are usually ignored. In semi-supervised learning, the training procedure is designed to extract useful structure from those unlabelled examples and combine it with the labelled set.
Common misconception: Semi-supervised learning and self-supervised learning are the same thing.
Correction: They overlap in practice but are not the same concept. Self-supervised learning creates training targets from the data itself, such as predicting a hidden part of an input. Semi-supervised learning uses some actual labels from outside that pretext task and a larger pool of unlabelled examples for the target task.

Quick check

Which situation best fits semi-supervised learning?

Go deeper

The simple version

You have a few examples with correct answers and many without answers. This approach uses both, so the many unanswered examples can still help the model learn patterns.

The technical version

Given labelled data {(x, y)} and additional unlabelled inputs {x}, semi-supervised learning estimates a predictor by combining a supervised objective with an unsupervised or consistency-based objective. Common mechanisms include pseudo-labelling, entropy minimisation, graph-based label propagation, consistency regularisation under augmentation, and generative modelling assumptions. These methods often rely on ideas such as the cluster assumption, smoothness assumption, or low-density separation: points close together, or in the same high-density region, are likely to share a label. Gains depend on task fit, calibration, data distribution, and careful use of validation and test sets.

Reviewed external resources

EncyclopaediaWeak supervision (opens in a new tab)Wikipedia
This Wikipedia page gives background on weak supervision, a related family of methods, and suits readers who want distinctions and terminology.
Intermediatelink checked 2026-07-31
CourseMachine Learning Crash Course (opens in a new tab)Google
This course offers practical machine-learning foundations and suits readers who want a structured introduction before placing semi-supervised methods in context.
Intermediatelink checked 2026-07-31
VideoStatQuest (opens in a new tab)StatQuest with Josh Starmer
This video source is useful for readers who prefer step-by-step spoken explanations and intuitive breakdowns of machine-learning ideas.
Starterlink checked 2026-07-31
VideoVideo explanations: Semi-supervised learning (opens in a new tab)YouTube search
This is a YouTube search results page, so it helps readers quickly find a range of video explanations of semi-supervised learning at different levels.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Semi-supervised learning (opens in a new tab)Reddit search
This is a Reddit search results page, so it suits readers who want to scan community questions, practical experiences and common confusions about semi-supervised learning.
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

Semi-supervised learning trains one predictor from both trusted labelled examples and additional unlabelled inputs. It can help when labels are scarce, but only when its assumptions and pseudo-labels hold up under evaluation.