Caret RF: poor results on new data with respect to partitioned data

i ran a RF through caret for a classification problem, according to the classical procedure:

  1. i used create DataPartition 2) i created a test dataset and a train dataset (from .csv file) 3) run with Caret a train utilizing "repeatedcv" as cross validation (n=10, repeats=5) 4) i got a good training model with fair accuracy (close to 0.75, similar sensitivity and specificity) 5) tested test data set (set up from createpartition) and got similar accuracy. Now, if a i use two DIFFERENT files for train and test data (that is to say that i manually separate data), that exactly form the same dataset partitioned at 1), then i train the model i obtain in the end very poor accuracy (around 0.5, in particular specificity drops to 0.3 and kappa goes down). How is it possible? Basically the only difference is that i held out the "test set" physically, but in data partition the system doesn't do the same? I saw the same behavior for another model i ran. below part of the script Than you very much in advance!!! best regards

I think that the script was left out.

Before you include it, you are most likely to get a good response by providing a reproducible example (= a reprex). The goal of a reprex is to make it as easy as possible for me to recreate your problem so that I can fix it: please help me help you!

If you've never heard of a reprex before, start by reading "What is a reprex", and follow the advice further down that page.

first of all thanks for your prompt reply. I tried to work with reprex, even though i was not able to include local .csv files.
i attach the lines from reprex, reffering to the case of SPLIT .csv files (train and test) that lead to very poor accuracy in testing data (and to the best case, where data partition is applied).If a basically create a partition of the whole SAME dataset and run the same rf i get very nice results on test data. i'm sorry but i'm not good at all in using reprex; i hope that preliminarly this new post may let you give me an insight.
best regards

CASE 1 - data split into to files
#> Carico il pacchetto richiesto: lattice
#> Carico il pacchetto richiesto: ggplot2
FMIB_10ind3_d_trainAll <- read_delim("FMIB_10ind3_v3_trainS.csv",
";", escape_double = FALSE, trim_ws = TRUE)
#> Error: 'FMIB_10ind3_v3_trainS.csv' does not exist in current working directory ('C:/Users/105034479/AppData/Local/Temp/RtmpKkfhcM/reprex31cc195b7785').
FMIB_10ind3_d_testAll <- read_delim("FMIB_10ind3_v3_testS.csv",
";", escape_double = FALSE, trim_ws = TRUE)
#> Error: 'FMIB_10ind3_v3_testS.csv' does not exist in current working directory ('C:/Users/105034479/AppData/Local/Temp/RtmpKkfhcM/reprex31cc195b7785').
dtrainA[["out10d"]] = factor(dtrainA[["out10d"]])
#> Error in factor(dtrainA[["out10d"]]): oggetto "dtrainA" non trovato
numFolds<-trainControl(method="repeatedcv", number =10, repeats=5)
#> Error in eval(expr, p): oggetto "dtrainA" non trovato
#> Error in predict(modelrf_10v2A, newdata = dtestA): oggetto "modelrf_10v2A" non trovato
#> Error in table(test_pred10v2A, dtestA$out10d): oggetto "test_pred10v2A" non trovato
#> Error in table(test_pred10v2A, dtestA$out10d): oggetto "test_pred10v2A" non trovato

CASE 2 data partitioned
FMIB_10ind3_v1 <- read_delim("FMIB_10ind3_v1.csv",
";", escape_double = FALSE, trim_ws = TRUE)
TDataIndex_v1 <- createDataPartition(y=FMIB_10ind3_v1$out10,p=0.7, list = FALSE)
trainingData10v1[["out10"]] = factor(trainingData10v1[["out10"]])
numFolds<-trainControl(method="repeatedcv", number=10,repeats=5)

This topic was automatically closed 21 days after the last reply. New replies are no longer allowed.