This chapter provides an in-depth exploration of the semi-supervised learning framework, beginning with foundational algorithms such as Expectation-Maximization (EM) and Classification Expectation-Maximization (CEM), which originated in the context of unsupervised learning. The discussion highlights how these algorithms underpin well-known models like K-means clustering, illustrating their role in discovering data structure without the need for labeled examples. EM is presented as a powerful iterative method for maximum likelihood estimation in probabilistic models with incomplete data, while CEM extends EM by integrating a classification step, with K-means emerging as a special case under specific assumptions. The chapter then introduces the core assumptions of semi-supervised learning and systematically examines three principal approaches: generative, discriminative, and graph-based methods. Generative approaches leverage both labeled and unlabeled data to model the underlying data distribution, improving classification performance when the distributional assumptions are met. Discriminative methods, such as self-training, focus on exploiting the geometry of the data and iteratively expand the labeled set by pseudo-labeling unlabeled instances. Graph-based approaches construct empirical graphs to capture the relationships between data points, propagating labels from a small set of labeled examples to a larger unlabeled population based on the assumption that similar instances are likely to share labels.