Text2Vec classification with caret - Naive Bayes warning message

Question

Please see the question listed here for more context.

I attempting to use a document term matrix, built using text2vec, to train a naive bayes (nb) model using the caret package. However, I get this warning message:

Warning message: In eval(xpr, envir = envir) : model fit failed for Fold01.Rep1: usekernel=FALSE, fL=0, adjust=1 Error in NaiveBayes.default(x, y, usekernel = FALSE, fL = param$fL, ...) : Zero variances for at least one class in variables:

Please help me to understand this message and what steps I need to take to avoid the model fitting from failing. I've a feeling that I need to remove more sparse terms from the DTM but I'm not sure.

Code to build the model:

    control <- trainControl(method="repeatedcv", number=10, repeats=3, savePredictions=TRUE, classProbs=TRUE)

    Train_PRDHA_String.df$Result <- ifelse(Train_PRDHA_String.df$Result == 1, "X", "Y")

    (warn=1)
    (warnings=2)

  t4 = Sys.time()
  svm_nb <- train(x = as.matrix(dtm_train), y = as.factor(Train_PRDHA_String.df$Result),
                  method = "nb",
                  trControl=control,
                  tuneLength = 5,
                  metric ="Accuracy")
print(difftime(Sys.time(), t4, units = 'sec'))

Code to build the Document Term Matrix (Text2Vec):

library(text2vec)
library(data.table)

#Define preprocessing function and tokenization fucntion
preproc_func = tolower
token_func = word_tokenizer

#Union both of the Text fields - learn vocab from both fields
union_txt = c(Train_PRDHA_String.df$MAKTX_Keyword, Train_PRDHA_String.df$PH_Level_04_Description_Keyword)

#Create an iterator over tokens with the itoken() function
it_train = itoken(union_txt, 
                  preprocessor = preproc_func, 
                  tokenizer = token_func, 
                  ids = Train_PRDHA_String.df$ID, 
                  progressbar = TRUE)

#Build Vocabulary
vocab = create_vocabulary(it_train)

vocab

#Dimensional Reduction
pruned_vocab = prune_vocabulary(vocab, 
                                term_count_min = 10, 
                                doc_proportion_max = 0.5,
                                doc_proportion_min = 0.001)
vectorizer = vocab_vectorizer(pruned_vocab)

#Start building a document-term matrix
#vectorizer = vocab_vectorizer(vocab)

#learn vocabulary from Train_PRDHA_String.df$MAKTX_Keyword
it1 = itoken(Train_PRDHA_String.df$MAKTX_Keyword, preproc_func, 
             token_func, ids = Train_PRDHA_String.df$ID)
dtm_train_1 = create_dtm(it1, vectorizer)

#learn vocabulary from Train_PRDHA_String.df$PH_Level_04_Description_Keyword
it2 = itoken(Train_PRDHA_String.df$PH_Level_04_Description_Keyword, preproc_func, 
             token_func, ids = Train_PRDHA_String.df$ID)
dtm_train_2 = create_dtm(it2, vectorizer)

#Combine dtm1 & dtm2 into a single matrix
dtm_train = cbind(dtm_train_1, dtm_train_2)

#Normalise
dtm_train = normalize(dtm_train, "l1")

dim(dtm_train)

Answer 1

It means, when these variables are resampled, they only have one unique value. You can use preProc = "zv" to get rid of the warning. It would help to get a small, reproducible example for these questions.