Which assumption states that unlabeled data naturally forms separable, locally dense clusters, implying nearby instances share the same label?

• A. Label Propagation
• B. Transductive semi-supervised learning
• C. Clustering Assumption
• D. Manifold Assumption

Answer: (C)

The clustering assumption says that data naturally form distinct, densely packed groups, and points that lie in the same group are likely to share the same label. Because of this, unlabeled points that fall inside a cluster can be inferred from the labeled points in that same cluster, since nearby instances tend to belong to the same class. It also implies that the decision boundary between classes should pass through the low-density regions between clusters, avoiding splitting a single cluster across classes.

This perspective is what underpins many semi-supervised methods that rely on the structure of the unlabeled data to propagate labels or to regularize the learning process, using the idea that clusters represent natural class membership. It differs from approaches grounded in graph-based label propagation, which emphasize smoothness across a network of data points, or from the manifold assumption, which emphasizes data lying on a lower-dimensional surface rather than explicit cluster density. It’s also a specific assumption about data structure, rather than a learning setting like transductive learning.