CODEWITHSUNDEEP
rsvm
user2955884
user2955884

Reputation: 574

Total accuracy in e1071::svm()

How is total accuracy actually calculated in e1071::svm()? It is not the same as calculated from the confusion table:

> x <- subset(iris, select = -Species)
> y <- iris$Species
> model <- svm(x, y,cross=10)
> model$tot.accuracy
[1] 96
> caret::confusionMatrix(y,model$fitted)
Confusion Matrix and Statistics

            Reference
Prediction   setosa versicolor virginica
  setosa         50          0         0
  versicolor      0         48         2
  virginica       0          2        48

Overall Statistics

               Accuracy : 0.9733          
  ...

Upvotes: 0

Views: 723

Answers (1)

Oliver
Oliver

Reputation: 8602

The total accuracy is the cross-validation accuracy obtained during the fitting procedure. For each split of the data, the model is fit and accuracy on the respective validation dataset is calculated.

The individual accuracy can be obtained from 

cv_accuracies <- model$accuracies
all.equal(mean(cv_accuracies), model$tot.accuracy)
[1] TRUE

Note that as the cross-validation splits the data into 10 different training and validation pairs, where the validation set is not used in the fitting procedure, this accuracy will be different compared to the accuracy of when predicting the final model to the entire training set.

sum(fitted(model) == y)/length(y)
[1] 0.97333333
mean(cv_accuracies)
[1] 96.66667

(note one is multiplied by 100 as standard)

Upvotes: 0

Related Questions