All MCQs

Distance-based methods are affected by the units and ranges of their features. Scaling gives variables a more comparable influence on neighborhood formation.

A very large neighborhood mixes distant observations and can ignore small but genuine local regions. The model may then underfit by favoring broad majority patterns.

One mislabeled or unusual training case can determine the prediction in its neighborhood. This low-bias, high-variance setting can overfit the training data.

k determines how many nearby labeled observations contribute to a prediction. Its value controls the locality and smoothness of the model.

Euclidean distance measures straight-line separation between feature vectors. Other distances may be used, but Euclidean distance is a common default for scaled numerical data.

k-NN predicts an outcome by using examples whose labels or response values are already known. Learning from labeled examples is the defining feature of supervised learning.

A negative coefficient means the fitted response decreases as that predictor increases, conditional on other model variables. The sign alone does not establish causation or practical importance.

Indicator or dummy variables encode category membership without imposing a false numerical distance. A reference category prevents redundant columns when an intercept is present.

Extrapolation assumes the fitted relationship continues into a region with no supporting training data. The true relationship may behave differently outside the observed range.

Highly correlated predictors make it difficult to separate their individual contributions. Coefficient estimates can become sensitive to small changes in the data.

Taking the square root reverses the squared units introduced by MSE. RMSE can therefore be compared directly with typical response values.

Mean squared error squares each residual and then takes their average. Large errors receive disproportionately greater weight because of the squaring.