A classic transductive semi‑supervised method.

A classic transductive semi‑supervised method.

• A. Self-training
• B. Graph regularization
• C. Label propagation
• D. Co-training

Answer: (C)

In transductive semi-supervised learning, you aim to assign labels to the specific unlabeled data points you have, using the relationships among all samples rather than building a general model for future data. Label propagation does exactly this: build a graph where each node is a data point and edge weights reflect similarity. The labeled points fix their labels, and the algorithm diffuses those labels across the graph so that nearby points end up with the same label. The result is a labeling for the unlabeled set without producing a model intended to generalize to new, unseen data. Self-training tends to be viewed as inductive, because it trains a classifier that will be used on new data beyond the initial unlabeled set. Co-training relies on multiple feature views and is more about leveraging those views to label data for retraining, rather than directly diffusing labels on a fixed graph. Graph regularization is related but label propagation is the quintessential, straightforward transductive method.

In transductive semi-supervised learning, you aim to assign labels to the specific unlabeled data points you have, using the relationships among all samples rather than building a general model for future data. Label propagation does exactly this: build a graph where each node is a data point and edge weights reflect similarity. The labeled points fix their labels, and the algorithm diffuses those labels across the graph so that nearby points end up with the same label. The result is a labeling for the unlabeled set without producing a model intended to generalize to new, unseen data.

Self-training tends to be viewed as inductive, because it trains a classifier that will be used on new data beyond the initial unlabeled set. Co-training relies on multiple feature views and is more about leveraging those views to label data for retraining, rather than directly diffusing labels on a fixed graph. Graph regularization is related but label propagation is the quintessential, straightforward transductive method.