All MCQs

The within-cluster sum of squares measures how compact the assigned groups are. K-means alternates updates to reduce this objective.

Each centroid represents the numerical center of one cluster in feature space. It is recalculated from the points currently assigned to that cluster.

k-means discovers groups using feature similarity without requiring known class labels. Its output is a partition of observations into clusters.

Validation estimates how different k values generalize without using the final test set. The selected k should balance local flexibility against oversmoothing.

k-NN represents the decision rule through stored data and local neighborhoods rather than a predetermined equation. Its effective complexity can grow with the amount of training data.

Distance weighting reflects the idea that very nearby cases may be more relevant to the query. Votes or response values are multiplied by weights that decrease with distance.

A naive neighbor search compares each new case with all training cases. Prediction can therefore become slow and memory-intensive for very large datasets.

k-NN does not estimate a compact parametric model during a conventional training phase. Neighbor searches and local prediction calculations occur when a query arrives.

The curse of dimensionality weakens the contrast between near and far points in high-dimensional spaces. Irrelevant features can therefore destroy meaningful neighborhood structure.

Equal class votes do not identify a unique majority prediction. The implementation must specify a policy such as distance weighting or a deterministic class priority.

k-NN regression combines the numerical outcomes of the selected neighbors, commonly using their mean. This produces a local estimate of the response.

After finding the k closest labeled cases, classification is usually based on their vote. The class receiving the most votes is assigned to the new case.