Variable importance plot for support vector machine with tidymodel framework is not working

Question

SVM does not have information within it about variable importance. So, I am using permutation of the variables to generate variable importance plot taking help from this answer, like

library(vip)
library(MASS)
library(tidymodels)

data(Boston, package = "MASS")
df <- Boston

#Split the data into train and test set
set.seed(7)
splits <- initial_split(df)
train <- training(splits)
test <- testing(splits)


#Preprocess with recipe
rec <- recipe(medv~.,data=train) %>%
  step_normalize(all_predictors()) 

svm_spec <- svm_rbf(margin = 0.0937, cost = 26.7, rbf_sigma = 0.0208) %>%
  set_engine("kernlab") %>%
  set_mode("regression")


#Putting into workflow
svr_fit <- workflow() %>%
  add_recipe(rec) %>%
  add_model(svm_spec) %>% 
  fit(data = train)

svr_fit %>%
  pull_workflow_fit() %>%
  vip(method = "permute", nsim = 5,
      target = "medv", metric = "rmse",
      pred_wrapper = kernlab::predict, train = train)

But it is returning me the following error:

Error in metric_fun(): ! estimate should be a numeric vector, not a numeric matrix.

At the same time pull_workflow_fit is deprecated. What should be used instead of pull_workflow_fit?

Answer 1

TL;DR pred_wrapper has to be updated to return a vector not a data.frame. The variables should not be normalized to get meaningful results. pull_workflow_fit is replaced with extract_fit_parsnip

library(vip)
#> 
#> Attaching package: 'vip'
#> The following object is masked from 'package:utils':
#> 
#>     vi
library(MASS)
library(tidymodels)

data(Boston, package = "MASS")
df <- Boston

# Split the data into train and test set
set.seed(7)
splits <- initial_split(df)
train <- training(splits)
test <- testing(splits)


# Preprocess with recipe
rec <- recipe(
  formula = medv ~ .,
  data = train
) 

svm_spec <- svm_rbf(margin = 0.0937, cost = 20, rbf_sigma = 0.0208) %>%
  set_engine("kernlab") %>%
  set_mode("regression")


# Putting into workflow
svr_fit <- workflow() %>%
  add_recipe(rec) %>%
  add_model(svm_spec) %>%
  fit(data = train)

svr_fit %>%
  extract_fit_parsnip() %>%
  vip(
    method = "permute", nsim = 5,
    target = "medv", metric = "rmse",
    geom = "point",
    pred_wrapper = function(object, newdata) as.vector(kernlab::predict(object, newdata)),
    train = train
  )

^{Created on 2023-12-30 with reprex v2.0.2}