[R] different randomForest performance for same data

[Uwe Ligges]

Häring wrote:
> Hello,
> 
> I came across a problem when building a randomForest model. Maybe someone can help me.
> I have a training- and a testdataset with a discrete response and ten predictors (numeric and factor variables). The two datasets are similar in terms of number of predictor, name of variables and datatype of variables (factor, numeric) except that only one predictor has got 20 levels in the training dataset and only 19 levels in the test dataset.
> I found that the model performance is different when train and test a model with the unchanged datasets on the one hand and after assigning the levels of the training dataset on the testdataset. I only assign the levels and do not change the dataset itself however the models perform different.
> Why???
> 
> Here is my code:
>> library(randomForest)
>> load("datasets.RData")  # import traindat and testdat
>> nlevels(traindat$predictor1)
> [1] 20
>> nlevels(testdat$predictor1)
> [1] 19
>> nrow(traindat)
> [1] 9838
>> nrow(testdat)
> [1] 3841
>> set.seed(10)
>> rf_orig <- randomForest(x=traindat[,-1], y=traindat[,1], xtest=testdat[,-1], ytest=testdat[,1],ntree=100)
>> data.frame(rf_orig$test$err.rate)[100,1]      # Error on test-dataset
> [1] 0.3082531
> 
> # assign the levels of the training dataset th the test dataset for predictor 1
>> levels(testdat$predictor1) <- levels(traindat$predictor1)  
>> nlevels(traindat$predictor1)
> [1] 20
>> nlevels(testdat$predictor1)
> [1] 20
>> nrow(traindat)
> [1] 9838
>> nrow(testdat)
> [1] 3841
>> set.seed(10)
>> rf_mod <- randomForest(x=traindat[,-1], y=traindat[,1], xtest=testdat[,-1], ytest=testdat[,1],ntree=100)
>> data.frame(rf_mod$test$err.rate)[100,1]       # Error on test-dataset
> [1] 0.4808644  # is different


Say testdat has 19 levels called L2, ..., L20 and traindat has 20 levels 
called L1, ..., L20.

After your call
  levels(testdat$predictor1) <- levels(traindat$predictor1)
You renamed L2 -> L1, L3 -> L2, ..., L20 -> L19 and invented a new level 
L20 that is unused.
Hence you confused all levels completely and given your ztrainikng is 
perfect, you will get an error rate of 100% in the end, because you 
renamed the levels in the testdata so that they do not fit to the 
traindata any more.

Uwe Ligges



> Cheers,
> TIM
> 
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.