Prediction for Manifold-to-Scalar Kernel Regression

Predicts under the fitted geometry and bandwidth without refitting or using prediction-batch statistics. Training predictors and responses are required.

Usage

# S3 method for class 'm2skreg'
predict(
  object,
  newdata,
  geometry = NULL,
  diagnostics = FALSE,
  block_size = 256L,
  ...
)

Arguments

object

A fitted m2skreg object.

newdata

A compatible riemdata object containing new predictors.

geometry

Usually NULL, which uses the fitted geometry. An explicit value must resolve to the same geometry. An old serialized object without geometry metadata requires an explicit value and produces a warning, because its original geometry cannot be inferred reliably.

diagnostics

Whether to return support diagnostics with predictions.

block_size

Positive integer limiting the number of new observations in each cross-distance calculation. It does not change the fitted model.

...

Reserved for future arguments; unknown arguments are rejected.

Value

By default, a numeric vector with one prediction per observation. With diagnostics=TRUE, a list containing prediction and a data frame diagnostics with effective_n, nearest_distance, and nearest_over_bandwidth. The last quantity may be infinite when the ratio is not representable even though the prediction remains finite.

See also

riem.m2skreg