Semi-Supervised Learning (SSL) bridges the gap between supervised and unsupervised learning by combining labeled and unlabeled data during training. Unlike pure supervised learning, SSL can leverage unlabeled data — but only when that data satisfies specific distributional assumptions: Smoothness, Cluster, Low-Density, and Manifold. The post explains the mathematical framing of SSL, distinguishes between inductive algorithms (which build a reusable model) and transductive algorithms (which propagate labels directly via graph-based methods), and honestly addresses the limitations: adding unlabeled data does not automatically improve performance. SSL is best treated as one tool in a researcher's toolkit, useful in domains where labeled data is scarce or expensive, such as image classification or drug discovery.
