Skip to content
ML Atlas

Semi-Supervised Learning

Learning from a few labeled examples plus many unlabeled ones, using the shape of the unlabeled data to spread the labels sensibly.

Also label propagation · pseudo-labeling · self-training · few labels

Role · kind
Core ideasParadigm
Level
Intermediate

You know the names of three kids in a big school photo. Kids standing in the same group are probably in the same class, so you can make good guesses about everyone else.

Where it fits

Learning ParadigmsLearning ParadigmsSupervised LearningSupervised LearningUnsupervised LearningUnsupervised LearningSemi-Supervised Learning
+All relationships as a list

Is a kind of / part of

See it on the full map

Try it

Same data, three ways to learn
Loading interactive widget…
TakeawayWith only a few labels, supervised learning draws a crude boundary. Semi-supervised learning lets the shape of the unlabeled data spread the labels, which works brilliantly when clusters match classes and fails when they do not.

How it works

Labels are often the expensive part of machine learning, while unlabeled data is plentiful. Semi-supervised learning sits between supervised and unsupervised learning: it uses a few labels for direction and lots of unlabeled data for shape.

The key assumption is that the data’s structure matches the classes: points in the same dense cluster or along the same curved band probably share a label. Two classic approaches:

  • Label propagation: connect each point to its nearest neighbors and let labels flow along those connections, like dye spreading through water, until every point has one.
  • Self-training (pseudo-labeling): train on the labeled points, label the unlabeled points the model is most confident about, add them to the training set, and repeat.

The widget shows the difference: with just a handful of labels, a supervised model draws a crude boundary, while label propagation follows the shape of the unlabeled data. When the clusters do not match the classes, the same trick confidently spreads errors.

Everyday use cases

  • Case 01

    Medical images

    Experts label a few hundred scans; thousands of unlabeled ones help the model learn what normal variation looks like.

  • Case 02

    Content moderation

    Label a small sample of posts, then let pseudo-labels cover the flood of unlabeled ones.

Strengths, weaknesses and when to use it

Strengths

  • Makes the most of scarce, expensive labels.

Weaknesses

  • If the data's clusters do not match the classes, spreading labels spreads mistakes.

When to use itWhen labels are expensive but unlabeled examples are plentiful.

Compare with

Check yourself

  1. Q01

    What assumption lets labels spread to unlabeled points?

    Show answer

    A: Points that are close together, in the same cluster, tend to share a label

Go deeper

Sources

  1. [1]van Engelen and Hoos, A survey on semi-supervised learning, Machine Learning (2020)
  2. [2]Zhou et al., Learning with Local and Global Consistency (2003)