Extract forward patterns, calibrated weights, or invariant maps

Usage

model_patterns(
  fit,
  type = c("forward", "weights", "conditional_info", "signal_sd"),
  dataset = NULL,
  ...
)

# S3 method for class 'pattern_fit'
model_importance(
  object,
  X_train = NULL,
  type = c("signal_sd", "conditional_info"),
  ...
)

# S3 method for class 'pattern_view'
model_importance(
  object,
  X_train = NULL,
  type = c("signal_sd", "conditional_info"),
  ...
)

# S3 method for class 'pattern_global_result'
model_importance(
  object,
  X_train = NULL,
  type = c("signal_sd", "conditional_info"),
  ...
)

Arguments

fit: A pattern_fit, pattern_view, or global result with a refit.
type: Quantity to extract.
dataset: Optional dataset for build_output_map. Global results supply their dataset automatically. Multibasis image maps are refused because aggregation across channels changes the estimand; extract vectors.
...: Reserved.
object: A fit, view, or global result.
X_train: Unused; maps are implied by the fitted model.

Value

A vector or matrix aligned to all input columns (screened columns are NA), or an image / list of component images when a dataset is supplied.

Details

Forward patterns and signal standard deviations are in original feature units. Weights map original, centred measurements to calibrated scores. Forward patterns and weights depend on the component basis. Signal SD and conditional information are coordinate invariant. Conditional information is in nats under the working Gaussian model for task scores, not empirical information about class labels. A feature with zero loading can have positive information through noise cancellation.