CODEWITHSUNDEEP
rr-caretroc
vorpalwabbit
vorpalwabbit

Reputation: 65

"ROC" metric not in result set

I'm trying to produce a random forest model with the caret package, using area under the ROC curve as the train metric, but I get the following warning:

Warning message:
In train.default(x = TrainData, y = TrainClasses, method = "rf",  :
  The metric "ROC" was not in the result set. Accuracy will be used instead.

Clearly this is not what I'm after, but I can't figure out where I'm going wrong.

Here is a reproducible example:

library(caret)
library(doParallel)
library(data.table)

cl <- makeCluster(detectCores() - 1) # I'm using 3 cores.
registerDoParallel(cl)

data(iris)
iris <- iris[iris$Species != 'virginica',] # to get two categories
TrainData <- as.data.table(iris[,1:4]) # My data is a data.table.
TrainClasses <- as.factor(as.character(iris[,5])) # to reset the levels to the two remaining flower types.

ctrl <- trainControl(method = 'oob',
                     classProbs = TRUE,
                     verboseIter = TRUE,
                     summaryFunction = twoClassSummary,
                     allowParallel = TRUE)
model.fit <- train(x = TrainData,
                   y = TrainClasses,
                   method = 'rf',
                   metric = 'ROC',
                   tuneLength = 3,
                   trControl = ctrl)

The result is the same if I don't create the parallel cluster and set allowParallel = FALSE.

In case it is of use, here's the result from a sessionInfo() call:

R version 3.2.2 (2015-08-14)
Platform: x86_64-w64-mingw32/x64 (64-bit)
Running under: Windows 7 x64 (build 7601) Service Pack 1

locale:
[1] LC_COLLATE=English_Australia.1252  LC_CTYPE=English_Australia.1252    LC_MONETARY=English_Australia.1252
[4] LC_NUMERIC=C                       LC_TIME=English_Australia.1252    

attached base packages:
[1] parallel  stats     graphics  grDevices utils     datasets  methods   base     

other attached packages:
[1] randomForest_4.6-10 data.table_1.9.6    doParallel_1.0.10   iterators_1.0.7     foreach_1.4.3      
[6] caret_6.0-52        ggplot2_1.0.1       lattice_0.20-33    

loaded via a namespace (and not attached):
 [1] Rcpp_0.12.1         compiler_3.2.2      nloptr_1.0.4        plyr_1.8.3          tools_3.2.2        
 [6] digest_0.6.8        lme4_1.1-8          nlme_3.1-121        gtable_0.1.2        mgcv_1.8-7         
[11] Matrix_1.2-2        brglm_0.5-9         SparseM_1.6         proto_0.3-10        BradleyTerry2_1.0-6
[16] stringr_1.0.0       gtools_3.5.0        stats4_3.2.2        grid_3.2.2          nnet_7.3-10        
[21] minqa_1.2.4         reshape2_1.4.1      car_2.0-26          magrittr_1.5        scales_0.3.0       
[26] codetools_0.2-14    MASS_7.3-44         splines_3.2.2       pbkrtest_0.4-2      colorspace_1.2-6   
[31] quantreg_5.11       stringi_0.5-5       munsell_0.4.2       chron_2.3-47

Thanks. Looking forward to getting this fixed!

Upvotes: 4

Views: 5777

Answers (1)

tchakravarty
tchakravarty

Reputation: 10954

You are correct. When you choose method = "oob", AUC-ROC is not one of the metrics that is returned.

You need to dig a little into the source code to figure out where the metrics are being computed. It is computed by method$oob which is called by oobTrainWorkflow on line 19, in turn called by train.default on line 258. method in your case is models$rf, where the object models is loaded from an external package file called models.RData: 

load(system.file("models", "models.RData", package = "caret"))

You can inspect the oob method for models$rf (which is the same as method): 

function(x) {
  out <- switch(x$type,
                regression =   c(sqrt(max(x$mse[length(x$mse)], 0)), x$rsq[length(x$rsq)]),
                classification =  c(1 - x$err.rate[x$ntree, "OOB"],
                                    e1071::classAgreement(x$confusion[,-dim(x$confusion)[2]])[["kappa"]]))
  names(out) <- if(x$type == "regression") c("RMSE", "Rsquared") else c("Accuracy", "Kappa")
  out
}

You can see that when classification RF is requested, only the accuracy and kappa metrics are computed.

You can tweak method$oob to use method$prob(mod$fit) and compute the AUC-ROC.

Upvotes: 3

Related Questions