Semi-supervised learning is a type of machine learning that trains a model on a small amount of labeled data together with a large amount of unlabeled data. It sits between supervised learning, which needs a label for every example, and unsupervised learning, which uses none. The approach is useful because labeling is often slow and expensive, since it needs people with expertise to review each example, while unlabeled data such as photos, audio and documents is cheap and abundant. By using both kinds of data, a semi-supervised model can often reach better accuracy than a model trained on the few labeled examples alone.
How semi-supervised learning works
The central idea is that unlabeled data contains information about the shape of the problem even though it has no answers. If two examples are very similar, they probably share a label, so the model can use the structure of the unlabeled data to spread what it learns from labeled points. In practice, the algorithm trains on the labeled examples to make an initial model, uses it together with the unlabeled data to refine its understanding of how examples are grouped, and then repeats. The unlabeled data guides where the boundary between classes should fall, ideally through regions where few examples lie.
Key assumptions
The approach works only when certain assumptions about the data hold.
- Smoothness assumption: examples that are close together in the input space are likely to have the same label.
- Cluster assumption: data tends to form clusters, and examples in the same cluster tend to share a class.
- Low-density separation: the boundary between classes should lie in regions with few data points.
- Manifold assumption: high-dimensional data often lies on a lower-dimensional surface, which makes distances more meaningful.
If these assumptions fail, for example when the unlabeled data comes from a different population, adding it can hurt performance instead of helping.
Common methods
Self-training, also called pseudo-labeling, trains a model on the labeled data, uses it to predict labels for the unlabeled data and adds the most confident predictions to the training set, and then retrains. Co-training uses two models that look at different views of the same data, such as the text of a web page and the text of links pointing to it, and each model labels examples for the other, a method proposed by Avrim Blum and Tom Mitchell in 1998. Consistency regularization encourages a model to give the same output for an example even after small changes, such as noise or cropping. Graph-based methods connect similar examples in a network and propagate labels along the connections. Generative approaches model how the data is produced and use that model to infer labels.
Applications of semi-supervised learning
Speech recognition benefits because recording audio is easy but transcribing it is costly. In medical imaging, specialists can label only a limited number of scans, so a model learns from those and from many unlabeled ones. Web page classification and text categorization use semi-supervised methods because the web is huge and labeled pages are few. Protein function prediction, fraud detection and image classification have also used the approach. Modern deep learning systems often combine semi-supervised ideas with self-supervised learning, first learning general features from unlabeled data and then fine-tuning with labels.
A worked example
Suppose a company wants to classify support tickets into topics. It has ten thousand tickets, but staff have time to label only five hundred. A model is trained on those five hundred and then predicts topics for the other nine and a half thousand tickets. Predictions with very high confidence are added to the training set as pseudo-labels, and the model is retrained on the larger set. This repeats for a few rounds, while the team checks a sample of pseudo-labeled tickets by hand to catch errors. The final model usually outperforms one trained on five hundred tickets alone, provided that the unlabeled tickets resemble those the model will see in use.
Advantages and risks
The approach lowers labeling costs, can improve accuracy and makes use of the large stores of unlabeled data that organizations already hold. The risks come from the unlabeled data. A model that makes early mistakes can reinforce them by training on its own wrong pseudo-labels, a problem called confirmation bias. If the unlabeled data differs from the labeled data, or contains classes absent from the labels, the model can learn misleading patterns. Teams therefore keep a clean, trusted validation set with real labels to measure progress and watch for degradation.
Semi-supervised learning in deep learning
Modern deep learning has produced stronger semi-supervised techniques. Methods such as FixMatch combine pseudo-labeling with consistency regularization. The model makes a prediction on a lightly altered image, keeps the prediction only if it is confident, and then trains the model to give the same answer for a heavily altered version of the same image. With only a few labeled examples per class, such methods have reached accuracy close to fully supervised models on image benchmarks. Language models use a similar logic when a small labeled set is used to fine-tune a model that was first pretrained on a large body of unlabeled text. These results explain why semi-supervised ideas remain central even as fully labeled datasets grow.
When to use semi-supervised learning
The approach makes most sense when labels are costly, the amount of unlabeled data is much larger than the labeled set and the two come from the same source. It is a poor fit when the labeled set is already large, when the unlabeled data is very different from the labeled data or when mistakes carry high stakes and cannot be allowed to compound. Teams often begin with a supervised baseline, add unlabeled data step by step, and compare results on a held-out labeled set. Active learning is a useful companion, in which the model asks people to label only the examples it finds most uncertain, so that scarce human effort goes where it helps most.
FAQs
How much labeled data does semi-supervised learning need? Often as few as one to ten percent of the dataset, though it depends on the task and the quality of the unlabeled data.
What is pseudo-labeling? It is a technique in which a model assigns its own predicted labels to unlabeled examples and then trains on them as if they were real labels.
How is it different from self-supervised learning? Semi-supervised learning uses some real labels. Self-supervised learning creates its training signal from the data itself and may use no human labels at all.
More in Machine Learning
Key facts
- Semi-supervised learning combines a small set of labeled examples with a much larger set of unlabeled examples during training.1
- Common methods include self-training with pseudo-labels, co-training, consistency regularization and graph-based label propagation.2
- It relies on assumptions such as smoothness and the cluster assumption, which say that nearby points tend to share a label.3
Related concepts
Quick checkWhat problem does semi-supervised learning mainly address?Show answer
The high cost of labeling data, by using a few labeled examples plus many cheap unlabeled ones to build a better model than the labeled data alone would allow.
Sources
- scikit-learn, Semi-supervised learning, https://scikit-learn.org/stable/modules/semi_supervised.html
- Wikipedia, Weak supervision, https://en.wikipedia.org/wiki/Weak_supervision
- Chapelle, Scholkopf and Zien, Semi-Supervised Learning, https://mitpress.mit.edu/9780262514125/semi-supervised-learning/
Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue
