How to use LIME to explain the best performing model?

Question

I am trying to perform model explainability for the best performing model using LIME for a classification problem. The y variable is whether a tumour is malignant or benign. Same question as (https://stackoverflow.com/questions/75867990/using-lime-in-r-to-explain-the-best-performing-model)

This is my attempt below:

models<-c("svmRadial","rf","knn")
results_table <- data.frame(models = models, stringsAsFactors = F)

for (i in models){
  model_train <- train(class~., data = training, method = i,
                       trControl= control, metric = "Accuracy")
  assign("fit", model_train)
  predictions <- predict(model_train, newdata = testing)
  table_mat<-table(testing$class, predictions)
  accuracy<-sum(diag(table_mat))/sum(table_mat)
  precision_ <-posPredValue(predictions, testing$class)
  recall_ <- sensitivity(predictions, testing$class)
  
  # put that in the results table
  results_table[results_table$models %in% i, "Accuracy"] <- accuracy
  results_table[results_table$models %in% i, "Precision"] <- precision_
  results_table[results_table$models %in% i, "Recall"] <- recall_
 
}

For this I got the following results, which are on results_table:

Model        Accuracy   Precision  Recall
svmRadial    0.9588235  0.9814815  0.954955
rf           0.9705882  0.9732143  0.981982
knn          0.9705882  0.9732143  0.981982

I have used LIME and it has worked (my attempt is below), but now I do not know which one of the models it is explaining. How do I know which model it is explaining? Is is explaining all three models or just the first model.

library(lime)
explainer_caret <- lime(training,  model_train)
 
explanation <- explain(testing[15:20, ], explainer_caret,
                       labels="malignant",
                       n_permutations=5,
                       dist_fun="manhattan",
                       kernel_width = 3,
                       n_features = 5)

Answer 1

It explains model_train. In your loop, you overwrite model_train three times (more precisely, you create it once and overwrite it twice). The last overwriting results in a model trained using method="knn", and that is what is explained.