Online Semi-supervised Learning
Online Semi-supervised Learning (Stream-based Learning with Partial Labels) · Also known as: stream-based semi-supervised learning, incremental semi-supervised learning, online SSL, semi-supervised online learning
Online semi-supervised learning combines the incremental, one-pass nature of online learning with the ability to exploit unlabeled data alongside sparse labeled observations. It is designed for settings where data arrives as a stream and obtaining labels for every instance is expensive or impractical — such as real-time classification of web content, sensor readings, or social media posts.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use online semi-supervised learning when data arrives as a continuous stream, labeling every instance is too costly, and the underlying distribution may drift over time. It suits text classification on live feeds, fraud detection, industrial sensor monitoring, and any domain with high-volume low-label-rate data. Do not use it when you have a static, fully labeled dataset — in that case standard batch semi-supervised or fully supervised methods are simpler and usually more accurate. Avoid it when the data stream is very short (fewer than a few hundred instances), as the model has little time to adapt before the stream ends.
Strengths & limitations
- Processes arbitrarily large data streams with bounded memory — no need to store the full dataset.
- Exploits unlabeled data to improve decision boundaries, reducing the labeling cost compared to purely supervised online learning.
- Continuously adapts to concept drift, making it suitable for non-stationary environments.
- Flexible labeling budget allows trading off annotation cost against prediction accuracy.
- Compatible with a wide range of base learners (linear models, kernel methods, neural networks).
- Pseudo-labels and unlabeled updates can propagate errors if the model's initial predictions are poor, leading to confirmation bias in early training.
- Hyperparameter choices (learning rate, regularization weight for unlabeled data, labeling budget) interact in complex ways and are harder to tune than in batch settings.
- Theoretical guarantees are weaker than in batch semi-supervised learning; performance depends heavily on the cluster or manifold assumption holding in the streaming data.
- Handling sudden concept drift while maintaining benefits from unlabeled data is non-trivial and requires careful drift-detection design.
Frequently asked
How is this different from regular online learning?
Regular online learning assumes every incoming instance has a label available immediately. Online semi-supervised learning relaxes this: most instances are unlabeled, and the model actively exploits their structure (manifold, cluster, or smoothness assumptions) to improve beyond what labeled data alone would support.
How is this different from batch semi-supervised learning?
Batch semi-supervised learning processes a fixed dataset in multiple passes and can use global graph structures over all data. Online SSL processes data in a single pass with bounded memory, building only local or approximate structures, which makes it applicable to streaming data but typically yields slightly lower accuracy than the batch counterpart.
What is the manifold assumption in this context?
The manifold assumption states that the decision boundary should lie in a low-density region and that nearby instances in the input space should receive similar predictions. In online SSL this is enforced locally: when an unlabeled instance arrives, the model is penalized if its prediction differs sharply from predictions on nearby recent instances.
How do I handle concept drift?
Use a drift detector (e.g., ADWIN or Page-Hinkley) to monitor prediction error or feature statistics. When drift is detected, increase the weight on newly labeled instances, reduce reliance on old pseudo-labels, and potentially reset or partially re-initialize the model.
What labeling budget strategy should I start with?
Start with uncertainty sampling — request a label when the model's confidence falls below a threshold. This concentrates the annotation budget on the instances most likely to improve the model. Combine it with a warm-up phase at the stream start where a higher fraction of labels is requested to bootstrap reliable initial predictions.
Sources
- Goldberg, A., Li, M., & Zhu, X. (2008). Online manifold regularization: A new learning setting and empirical study. In Proceedings of the European Conference on Machine Learning and Knowledge Discovery in Databases (ECML PKDD), pp. 393–407. Springer. link ↗
- Semi-supervised learning. Wikipedia. link ↗
How to cite this page
ScholarGate. (2026, June 3). Online Semi-supervised Learning (Stream-based Learning with Partial Labels). ScholarGate. https://scholargate.app/en/machine-learning/online-semi-supervised-learning
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Active LearningMachine learning↔ compare
- Label PropagationMachine learning↔ compare
- Online LearningMachine learning↔ compare
- Self-supervised LearningMachine learning↔ compare
- Semi-supervised LearningMachine learning↔ compare
- Transfer LearningMachine learning↔ compare