CODEWITHSUNDEEP
rr-caretroc
Andres Mora
Andres Mora

Reputation: 1106

ROC in Multiclass kNN

Im trying to run some ROC analysis on a multiclass knn model and dataset

so far i have this code for the kNN model. It works well.
X_train_new is a dataset with 131 numeric variables (columns) and 7210 observations.
Y_train is the outcome variable which i have as factor. its a dataset with only 1 column (activity) and 7210 observations (there are 6 possible factors)

ctrl <- trainControl(method  = "cv",
                     number  = 10)

model2 <-    train(X_train_new,
                   Y_train$activity,
                   method     = "knn",
                   tuneGrid   = expand.grid(k = 5),
                   trControl  = ctrl,
                   metric     = "Accuracy"
)

X_test_new is a dataset with 131 numeric variables (columns) and 3089 observations.
Y_test is the outcome variable which i have as factor. its a dataset with only 1 column and 3089 observations (there are 6 possible factors)

I run the predict function

knnPredict_test <- predict(model2 , newdata = X_test_new )

I would like to do some ROC analysis on each class vs all. Im trying

a = multiclass.roc ( Y_test$activity, knnPredict_test )

knnPredict_test is a vector with predicted classes:

knnPredict_test <- predict(model2 ,newdata = X_test_new )
> length(knnPredict_test)
[1] 3089
> glimpse(knnPredict_test)
 Factor w/ 6 levels "laying","sitting",..: 2 1 5 1 3 2 4 5 3 2 ...

This is the error im getting

Error in roc.default(response, predictor, levels = X, percent = percent,  :   
Predictor must be numeric or ordered.

Upvotes: 2

Views: 1006

Answers (1)

Paul
Paul

Reputation: 9087

To get the ROC, you need a numeric prediction. However, by default predict will give you the predicted classes. Use type = "prob".

Here is a reproducable example which has the same error.

library(caret)

knnFit <- train(
  Species ~ .,
  data = iris,
  method = "knn"
)

predictions_bad <- predict(knnFit)

pROC::multiclass.roc(iris$Species, predictions_bad)
#> Error in roc.default(response, predictor, levels = X, percent = percent,  : 
#>   Predictor must be numeric or ordered.

Using type = "prob" fixes the error.

predictions_good <- predict(knnFit, type = "prob")

pROC::multiclass.roc(iris$Species, predictions_good)
#> Call:
#> multiclass.roc.default(response = iris$Species, predictor = predictions_good)
#> 
#> Data: multivariate predictor predictions_good with 3 levels of iris$Species: setosa, versicolor, virginica.
#> Multi-class area under the curve: 0.9981

Upvotes: 2

Related Questions