A small subset of reviews has human-verified sentiment labels, alongside many unlabeled reviews. Compare ways to train sentiment classification for new reviews.
Distinguish training only on the labeled subset from training with both labeled and unlabeled data. Keep a separate human-labeled evaluation dataset out of training.
Choose TWO appropriate descriptions based on the data actually used for learning.
Check the data actually used for training, not merely available data. Labeled-only training is supervised; combining labeled and unlabeled data can support semi-supervised learning. Handle evaluation truth and pseudo-label quality separately.
Detailed explanation
Pseudo-labels are model estimates and can be wrong. Check label provenance and quality rather than treating them unconditionally as evaluation truth.
This combines labeled and unlabeled data. Check whether both are actually used in learning rather than merely available.
Reinforcement learning concerns signals such as actions and rewards. Data volume alone does not distinguish semi-supervised, unsupervised and reinforcement learning.
Unlabeled examples have no verified answers for comparison. Use an appropriate labeled evaluation dataset and identify the reference outcome.
The model learns from input examples paired with verified labels. The existence of other unused unlabeled data does not change the learning method.
Check official documentation
Compare the following points with the official documentation.
Compare AWS supervised and semi-supervised definitions and identify each proposal use of labeled data, unlabeled data and pseudo-labels. Check that verified evaluation answers remain outside training.
Official references
What to confirm
Distinguish labeled-only training from training that also uses unlabeled data. Unlabeled examples or pseudo-labels alone do not establish ordinary classification accuracy against verified evaluation answers.
Key points
- Actual training data
- Semi-supervised learning
- Pseudo-labels
- Evaluation data
Notes
- Environment: Official AIF-C01 Task 1.1 and AI/ML documentation review; no model training or production changes are performed.
- Compare the definitions with current AIF-C01 Task 1.1 and the linked primary sources.
- A conceptual example does not establish measured model accuracy or authorize external operations.
Foundation review
Separate availability from use
Possessing unlabeled examples does not turn a classifier trained only on labeled examples into semi-supervised learning. Identify the data actually used by each training proposal.
Check label provenance
Human-verified labels, unlabeled inputs and model-estimated pseudo-labels have different roles and quality. Evaluate errors and bias when using pseudo-labels.
Keep evaluation answers separate from learning
Classification accuracy needs reference answers to compare with predictions. Unlabeled volume or training performance alone does not establish quality on unseen data.