first time poster here, so apologies for rookie errors
I am using the caret package in R for classification. I am fitting some models (GBM, linear SVM, NB, LDA) using repeated 10-fold cross validation over a training set. Using a custom trainControl, caret even gives me a whole range of model performance metrics like ROC, Spec/sens, Kappa, Accuracy over the test folds. That really is fantastic. There is just one more metric I would love to have: some measure of model calibration.
I noticed that there is a function within caret that can create a calibration plot to estimate the consistency of model performance across portions of your data. Is it possible to have caret compute this for each test-fold during the cross-validated model building? Or can it only be applied to some new held out data that we are making predictions on?
For some context, at the moment I have something like this:
fitControl <- trainControl(method = "repeatedcv", repeats=2, number = 10, classProbs = TRUE, summaryFunction = custom.summary)
gbmGrid <- expand.grid(.interaction.depth = c(1,2,3),.n.trees = seq(100,800,by=100),.shrinkage = c(0.01))
gbmModel <- train(y= train_target, x = data.frame(t_train_predictors),
method = "gbm",
trControl = fitControl,
tuneGrid = gbmGrid,
verbose = FALSE)
If it helps, I am using ~25 numeric predictors and N=2,200, predicting a two-class factor.
Many thanks in advance for any help/advice. Adam
The calibration
function takes whatever data that you give it. You can get the resampled values from the train
sub-object pred
:
> set.seed(1)
> dat <- twoClassSim(2000)
>
> set.seed(2)
> mod <- train(Class ~ ., data = dat,
+ method = "lda",
+ trControl = trainControl(savePredictions = TRUE,
+ classProbs = TRUE))
>
> str(mod$pred)
'data.frame': 18413 obs. of 7 variables:
$ pred : Factor w/ 2 levels "Class1","Class2": 1 2 2 1 1 2 1 1 2 1 ...
$ obs : Factor w/ 2 levels "Class1","Class2": 1 2 2 1 1 2 1 1 2 2 ...
$ Class1 : num 0.631 0.018 0.138 0.686 0.926 ...
$ Class2 : num 0.369 0.982 0.8616 0.3139 0.0744 ...
$ rowIndex : int 1 3 4 10 12 13 18 22 25 27 ...
$ parameter: Factor w/ 1 level "none": 1 1 1 1 1 1 1 1 1 1 ...
$ Resample : chr "Resample01" "Resample01" "Resample01" "Resample01" ...
Then you could use:
> cal <- calibration(obs ~ Class1, data = mod$pred)
> xyplot(cal)
Just keep in mind that, with many resampling methods, a single training set instance will be held-out multiple times:
> table(table(mod$pred$rowIndex))
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
2 11 30 77 135 209 332 314 307 231 185 93 48 16 6 4
You could average the class probabilities per rowIndex
if you like.
Max