CODEWITHSUNDEEP
rmachine-learningr-caretrpart
chrisjacques
chrisjacques

Reputation: 646

Optimising caret for sensitivity still seems to optimise for ROC

I'm trying to maximise sensitivity in my model selection in caret using rpart. To this end, I tried to replicate the method given here (scroll down to the example with the user-defined function FourStat) caret's github page

# create own function so we can use "sensitivity" as our metric to maximise:
Sensitivity.fc <- function (data, lev = levels(data$obs), model = NULL) {
    out <- c(twoClassSummary(data, lev = levels(data$obs), model = NULL))
    c(out, Sensitivity = out["Sens"])
}

rpart_caret_fit <- train(outcome~pred1+pred2+pred3+pred4,
    na.action = na.pass,
    method = "rpart", 
    control=rpart.control(maxdepth = 6),
    tuneLength = 20, 
    # maximise sensitivity
    metric = "Sensitivity", 
    maximize = TRUE,
    trControl = trainControl(classProbs = TRUE,
    summaryFunction = Sensitivity.fc))

However when I get the summary with 

rpart_caret_fit

it indicates that it still used the ROC criterion to select the final model:

CART 

678282 samples
 4 predictor
 2 classes: 'yes', 'no' 

No pre-processing
Resampling: Bootstrapped (25 reps) 
Summary of sample sizes: 678282, 678282, 678282, 678282, 678282, 678282, ... 
Resampling results across tuning parameters:

cp              ROC        Sens       Spec       Sensitivity.Sens
0.000001909738  0.7259486  0.4123547  0.8227382  0.4123547       
0.000002864607  0.7259486  0.4123547  0.8227382  0.4123547       
0.000005729214  0.7259489  0.4123622  0.8227353  0.4123622       
0.000006684083  0.7258036  0.4123614  0.8227379  0.4123614       
0.000007638953  0.7258031  0.4123576  0.8227398  0.4123576       
0.000009548691  0.7258028  0.4123539  0.8227416  0.4123539       
0.000010694534  0.7257553  0.4123589  0.8227332  0.4123589       
0.000015277905  0.7257313  0.4123614  0.8227290  0.4123614       
0.000032465548  0.7253456  0.4112838  0.8234272  0.4112838       
0.000038194763  0.7252966  0.4112912  0.8234196  0.4112912       
0.000076389525  0.7248774  0.4102792  0.8240339  0.4102792       
0.000164237480  0.7244847  0.4093688  0.8246372  0.4093688       
0.000194793290  0.7241532  0.4086596  0.8250930  0.4086596       
0.000310650737  0.7237546  0.4087379  0.8250393  0.4087379       
0.001625187154  0.7233805  0.4006570  0.8295729  0.4006570       
0.001726403276  0.7233225  0.3983850  0.8308874  0.3983850       
0.002173282000  0.7230906  0.3915758  0.8348320  0.3915758       
0.002237258227  0.7230906  0.3915758  0.8348320  0.3915758       
0.006140444689  0.7173854  0.4897494  0.7695558  0.4897494       
0.055330843035  0.5730987  0.2710906  0.8545549  0.2710906       

ROC was used to select the optimal model using the largest value.
The final value used for the model was cp = 0.000005729214.

How can I override the ROC selection method?

Upvotes: 4

Views: 1745

Answers (1)

missuse
missuse

Reputation: 19756

You over-complicated things.

Two class summary already contains Sensitivity as output. The column name "Sens". It is enough to specify:

metric = "Sens" to train and summaryFunction = twoClassSummary to trainControl

Full example:

library(caret)
library(mlbench)
data(Sonar)

rpart_caret_fit <- train(Class~., 
                         data = Sonar,
                         method = "rpart", 
                         tuneLength = 20, 
                         metric = "Sens", 
                         maximize = TRUE,
                         trControl = trainControl(classProbs = TRUE,
                                                  method = "cv",
                                                  number = 5,
                                                  summaryFunction = twoClassSummary))

rpart_caret_fit
CART 

208 samples
 60 predictor
  2 classes: 'M', 'R' 

No pre-processing
Resampling: Cross-Validated (5 fold) 
Summary of sample sizes: 167, 166, 166, 166, 167 
Resampling results across tuning parameters:

  cp         ROC        Sens       Spec     
  0.0000000  0.7088298  0.7023715  0.7210526
  0.0255019  0.7075400  0.7292490  0.6684211
  0.0510038  0.7105388  0.7758893  0.6405263
  0.0765057  0.6904202  0.7841897  0.6294737
  0.1020076  0.7104681  0.8114625  0.6094737
  0.1275095  0.7104681  0.8114625  0.6094737
  0.1530114  0.7104681  0.8114625  0.6094737
  0.1785133  0.7104681  0.8114625  0.6094737
  0.2040152  0.7104681  0.8114625  0.6094737
  0.2295171  0.7104681  0.8114625  0.6094737
  0.2550190  0.7104681  0.8114625  0.6094737
  0.2805209  0.7104681  0.8114625  0.6094737
  0.3060228  0.7104681  0.8114625  0.6094737
  0.3315247  0.7104681  0.8114625  0.6094737
  0.3570266  0.7104681  0.8114625  0.6094737
  0.3825285  0.7104681  0.8114625  0.6094737
  0.4080304  0.7104681  0.8114625  0.6094737
  0.4335323  0.7104681  0.8114625  0.6094737
  0.4590342  0.6500135  0.8205534  0.4794737
  0.4845361  0.6500135  0.8205534  0.4794737

Sens was used to select the optimal model using the largest value.
The final value used for the model was cp = 0.4845361.

Additionally I do not think you can specify control = rpart.control(maxdepth = 6) to caret train. This is not correct - caret passes any parameters forward using .... So you can pass pretty much any argument.

If you are looking to write you own summary functions here is an example on the "Sens":

Sensitivity.fc <- function (data, lev = NULL, model = NULL) { #every summary function takes these three arguments
  obs <- data[, "obs"] #these are the real values - always in column name "obs" in data
  cls <- levels(obs) #there are the levels - you can also pass this to lev argument 
  probs <- data[, cls[2]] #these are the probabilities for the 2nd class - useful only if prob = TRUE
  class <- as.factor(ifelse(probs > 0.5, cls[2], cls[1])) #calculate the classes based on some probability treshold
  Sensitivity <- caret::sensitivity(class, obs) #do the calculation - I was lazy so I used a built in function to do it for me
  names(Sensitivity) <- "Sens" #the name of the output
  Sensitivity
}

and now:

rpart_caret_fit <- train(Class~., 
                         data = Sonar,
                         method = "rpart", 
                         tuneLength = 20, 
                         metric = "Sens", #because of this line: names(Sensitivity) <- "Sens" 
                         maximize = TRUE,
                         trControl = trainControl(classProbs = TRUE,
                                                  method = "cv",
                                                  number = 5,
                                                  summaryFunction = Sensitivity.fc))

Lets check if both produce the same results:

set.seed(1)
fit_sens <- train(Class~., 
                  data = Sonar,
                  method = "rpart", 
                  tuneLength = 20, 
                  metric = "Sens", 
                  maximize = TRUE,
                  trControl = trainControl(classProbs = TRUE,
                                           method = "cv",
                                           number = 5,
                                           summaryFunction = Sensitivity.fc))

set.seed(1)
fit_sens2 <- train(Class~., 
                   data = Sonar,
                   method = "rpart", 
                   tuneLength = 20, 
                   metric = "Sens", 
                   maximize = TRUE,
                   trControl = trainControl(classProbs = TRUE,
                                            method = "cv",
                                            number = 5,
                                            summaryFunction = twoClassSummary))

all.equal(fit_sens$results[c("cp", "Sens")],
          fit_sens2$results[c("cp", "Sens")])  

TRUE

all.equal(fit_sens$bestTune,
          fit_sens2$bestTune)
TRUE

Upvotes: 9

Related Questions