Học bán giám sát

A training collection may contain some known target labels and many examples without them. Semi-supervised methods aim to use information in both parts. More unlabeled examples do not automatically improve a classifier: their distribution and the method's assumptions matter. Self-training starts from a supervised classifier and selects model predictions on unlabeled examples as pseudo-labels for further fitting. A confidence threshold or a ranked selection rule can determine which guesses are added. Incorrect guesses can still be reinforced by training, including when the model is confidently wrong. Consistency regularization encourages predictions to agree across chosen perturbations or augmentations intended to preserve the relevant class. It should not be generalized into a rule that every small-looking edit preserves meaning. Cluster-based reasoning motivates some approaches by assuming that grouped observations are likely to share a label, but this is an assumption to examine rather than a fact about every dataset or method. FixMatch combines a particular pseudo-labeling procedure with consistency. A confident prediction on a weakly augmented image supplies a pseudo-label, and a strongly augmented view is trained toward that label. The augmentation choices and threshold define the procedure, and its benchmark findings remain specific to evaluated conditions. Label propagation uses another mechanism: a similarity graph over labeled and unlabeled observations supports inferred label information for related points. Its usefulness depends on whether the graph reflects relevant class structure. An implementation detail matters when using scikit-learn's semi-supervised interface: unlabeled entries in the target array use the integer marker minus one, distinct from actual class labels.