There is a smooth and continuous boundary separating classes. If two points are 'close' in the feature space, their labels should be the same.

There is a smooth and continuous boundary separating classes. If two points are 'close' in the feature space, their labels should be the same.

• A. Smoothness (Continuity) Assumption
• B. Clustering Assumption
• C. Manifold Assumption
• D. High-dimensional data

Answer: (A)

The main idea here is smoothness in how the label changes across the feature space. If the boundary separating classes is smooth and continuous, then moving a little in input space shouldn’t flip the label. In other words, nearby points are likely to belong to the same class because the decision function changes gradually rather than abruptly. This locality-based view underpins many learning methods that rely on nearby examples having similar labels, such as k-nearest neighbors and kernel-based approaches, as well as semi-supervised techniques that propagate label information through regions of the space that appear homogeneous. The clustering assumption is related but distinct: it says points within the same cluster share the same label, but clusters can be irregular and don’t inherently guarantee label consistency right at the boundary or across all nearby points. The manifold assumption focuses on the geometry of the data—data lie on a lower-dimensional surface within a higher-dimensional space—without directly asserting that close points must have the same label. High-dimensional data describes the space’s nature rather than a rule about how labels behave with proximity.

The main idea here is smoothness in how the label changes across the feature space. If the boundary separating classes is smooth and continuous, then moving a little in input space shouldn’t flip the label. In other words, nearby points are likely to belong to the same class because the decision function changes gradually rather than abruptly. This locality-based view underpins many learning methods that rely on nearby examples having similar labels, such as k-nearest neighbors and kernel-based approaches, as well as semi-supervised techniques that propagate label information through regions of the space that appear homogeneous.

The clustering assumption is related but distinct: it says points within the same cluster share the same label, but clusters can be irregular and don’t inherently guarantee label consistency right at the boundary or across all nearby points. The manifold assumption focuses on the geometry of the data—data lie on a lower-dimensional surface within a higher-dimensional space—without directly asserting that close points must have the same label. High-dimensional data describes the space’s nature rather than a rule about how labels behave with proximity.