Semi-Supervised Learning combines labeled data with a larger set of unlabeled data to improve model performance. Methods include label propagation, consistency regularization, and self-training. It is especially useful when labeled data is scarce but unlabeled data is abundant.