CODEWITHSUNDEEP
rtidymodelsvip
itsMeInMiami
itsMeInMiami

Reputation: 2763

How can I get Variable Importance Plot for a categorical SVM in tidymodels

I would like to get VIP info on a SVM model that is doing classification. I found this useful post Variable importance plot for support vector machine with tidymodel framework is not working which shows how to get the plot for a regression model and I tried to tweak it.

Unfortunately it throws an error. Can anyone please tell me what I am doing wrong?

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

data(Boston, package = "MASS")

# Make a classificaiton outcome
df <- Boston |> 
  mutate(is_big = factor(if_else(medv > 22, 1, 0)))

# 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 = is_big ~ .,
  data = train
) 

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


# 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 = "is_big", metric = "roc_auc", event_level = "second",
    geom = "point",
    pred_wrapper = 
      function(object, newdata) as.vector(kernlab::predict(object, newdata)),
    train = train
  )
#> Error in `fun()`:
#> ! `estimate` should be a numeric vector, not a character vector.

Created on 2024-08-04 with reprex v2.1.1

Upvotes: 1

Views: 122

Answers (1)

bgreenwell
bgreenwell

Reputation: 413

I think there's a bug in the vip() function for tidymodels workflows, you can work around it by calling vi() or vi_permute() directly or by calling vip() on the raw underlying model fit. Here's a working version using both approaches:

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

data(Boston, package = "MASS")

# Make a classificaiton outcome
df <- Boston |> 
  mutate(is_big = factor(if_else(medv > 22, 1, 0)))

# 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 = is_big ~ .,
  data = train
) 

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


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

# Extract the raw underlying fit
original_fit <- workflows::extract_fit_engine(svr_fit)

# Prediction wrapper should return a vector of probabilities for the second class
pfun <- function(object, newdata) {
  kernlab::predict(object, newdata, type = "probabilities")[, 2L]
}

# Now this should work
original_fit %>%
  vip(
    method = "permute",
    nsim = 5,
    target = "is_big", metric = "roc_auc", event_level = "second",
    pred_wrapper = pfun,
    train = train
  )

# Alternatively, you can define a prediction wrapper for the workflow object 
# directly; vip() seems to be bugged with tidymodels workflows
svr_fit %>%
  vi(
    method = "permute",
    nsim = 5,
    target = "is_big", metric = "roc_auc", event_level = "second",
    pred_wrapper = function(object, newdata) predict(object, newdata, type = "prob")[[".pred_1"]],
    train = train
  )

Careful though! Your example includes leakage since your binary outcome is a direct function of medv which is also included as an input; hence the large importance score for the latter.

Upvotes: 2

Related Questions