vivi: vivi

In vivid: Variable Importance and Variable Interaction Displays

vivi

Description

Creates a matrix displaying variable importance on the diagonal and variable interaction on the off-diagonal.

Usage

vivi(
  data,
  fit,
  response,
  gridSize = 50,
  importanceType = "agnostic",
  nmax = 500,
  reorder = TRUE,
  class = 1,
  predictFun = NULL,
  normalized = FALSE,
  numPerm = 4,
  showVimpError = FALSE
)

Arguments

data	Data frame used for fit.
fit	A supervised machine learning model, which understands condvis2::CVpredict
response	The name of the response for the fit.
gridSize	The size of the grid for evaluating the predictions.
importanceType	Used to select the importance metric. By default, an agnostic importance measure is used. If an embedded metric is available, then setting this argument to the importance metric will use the selected importance values in the vivid-matrix. Please refer to the examples given for illustration. Alternatively, set to equal "agnostic" (the default) to override embedded importance measures and return agnostic importance values.
nmax	Maximum number of data rows to consider. Default is 500. Use all rows if NULL.
reorder	If TRUE (default) uses DendSer to reorder the matrix of interactions and variable importances.
class	Category for classification, a factor level, or a number indicating which factor level.
predictFun	Function of (fit, data) to extract numeric predictions from fit. Uses condvis2::CVpredict by default, which works for many fit classes.
normalized	Should Friedman's H-statistic be normalized or not. Default is FALSE.
numPerm	Number of permutations to perform for agnostic importance. Default is 4.
showVimpError	Logical. If TRUE, and numPerm > 1 then a tibble containing the variable names, their importance values, and the standard error for each importance is printed to the console.

Details

If the argument importanceType = 'agnostic', then an agnostic permutation importance (1) is calculated. Friedman's H statistic (2) is used for measuring the interactions. This measure is based on partial dependence curves and relates the interaction strength of a pair of variables to the total effect strength of that variable pair.

Value

A matrix of interaction values, with importance on the diagonal.

References

1: Fisher A., Rudin C., Dominici F. (2018). All Models are Wrong but many are Useful: Variable Importance for Black-Box, Proprietary, or Misspecified Prediction Models, using Model Class Reliance. Arxiv.

2: Friedman, J. H. and Popescu, B. E. (2008). “Predictive learning via rule ensembles.” The Annals of Applied Statistics. JSTOR, 916–54.

Examples

aq <- na.omit(airquality)
f <- lm(Ozone ~ ., data = aq)
m <- vivi(fit = f, data = aq, response = "Ozone") # as expected all interactions are zero
viviHeatmap(m)

# Select importance metric
library(randomForest)
rf1 <- randomForest(Ozone~., data = aq, importance = TRUE)
m2 <- vivi(fit = rf1, data = aq, response = 'Ozone',
           importanceType = '%IncMSE') # select %IncMSE as the importance measure
viviHeatmap(m2)


library(ranger)
rf <- ranger(Species ~ ., data = iris, importance = "impurity", probability = TRUE)
vivi(fit = rf, data = iris, response = "Species") # returns agnostic importance
vivi(fit = rf, data = iris, response = "Species",
     importanceType = "impurity") # returns selected 'impurity' importance.

vivid documentation built on July 26, 2023, 5:22 p.m.