Error when I try to predict class probabilities in R - caret - r

Question

I've build a model using caret. When the training was completed I got the following warning:
Warning message:
In train.default(x, y, weights = w, ...) :
At least one of the class levels are not valid R variables names; This may cause errors if class probabilities are generated because the variables names will be converted to: X0, X1
The names of the variables are:
str(train)
'data.frame': 7395 obs. of 30 variables:
$ alchemy_category : Factor w/ 13 levels "arts_entertainment",..: 2 8 6 6 11 6 1 6 3 8 ...
$ alchemy_category_score : num 3737 2052 4801 3816 3179 ...
$ avglinksize : num 2.06 3.68 2.38 1.54 2.68 ...
$ commonlinkratio_1 : num 0.676 0.508 0.562 0.4 0.5 ...
$ commonlinkratio_2 : num 0.206 0.289 0.322 0.1 0.222 ...
$ commonlinkratio_3 : num 0.0471 0.2139 0.1202 0.0167 0.1235 ...
$ commonlinkratio_4 : num 0.0235 0.1444 0.0426 0 0.0432 ...
$ compression_ratio : num 0.444 0.469 0.525 0.481 0.446 ...
$ embed_ratio : num 0 0 0 0 0 0 0 0 0 0 ...
$ frameTagRatio : num 0.0908 0.0987 0.0724 0.0959 0.0249 ...
$ hasDomainLink : Factor w/ 2 levels "0","1": 1 1 1 1 1 1 1 1 1 1 ...
$ html_ratio : num 0.246 0.203 0.226 0.266 0.229 ...
$ image_ratio : num 0.00388 0.08865 0.12054 0.03534 0.05047 ...
$ is_news : Factor w/ 2 levels "0","1": 2 2 2 2 2 1 2 1 2 1 ...
$ lengthyLinkDomain : Factor w/ 2 levels "0","1": 2 2 2 1 2 1 1 1 1 2 ...
$ linkwordscore : num 24 40 55 24 14 12 21 5 17 14 ...
$ news_front_page : Factor w/ 2 levels "0","1": 1 1 1 1 1 1 1 1 1 1 ...
$ non_markup_alphanum_characters: num 5424 4973 2240 2737 12032 ...
$ numberOfLinks : num 170 187 258 120 162 55 93 132 194 326 ...
$ numwords_in_url : num 8 9 11 5 10 3 3 4 7 4 ...
$ parametrizedLinkRatio : num 0.1529 0.1818 0.1667 0.0417 0.0988 ...
$ spelling_errors_ratio : num 0.0791 0.1254 0.0576 0.1009 0.0826 ...
$ label : Factor w/ 2 levels "0","1": 1 2 2 2 1 1 2 1 2 2 ...
$ isVideo : Factor w/ 2 levels "0","1": 2 2 2 2 2 2 2 2 1 1 ...
$ isFashion : Factor w/ 2 levels "0","1": 1 1 1 1 2 1 2 1 2 1 ...
$ isFood : Factor w/ 2 levels "0","1": 2 2 2 2 2 2 2 2 2 2 ...
$ hasComments : Factor w/ 2 levels "0","1": 1 2 2 2 2 1 2 2 1 2 ...
$ hasGoogleAnalytics : Factor w/ 2 levels "0","1": 1 1 1 1 2 1 2 2 2 1 ...
$ hasInlineCSS : Factor w/ 2 levels "0","1": 1 2 2 2 1 1 2 1 2 2 ...
$ noOfMetaTags : num 10 12 6 10 13 2 6 6 9 5 ...
My code is the following:
ctrl <- trainControl(method = "CV",
number=10,
classProbs = TRUE,
allowParallel = TRUE,
summaryFunction = twoClassSummary)
set.seed(476)
rfFit <- train(formula,
data=train,
method = "rf",
tuneGrid = expand.grid(.mtry = seq(4,20,by=2)),
ntrees=1000,
importance = TRUE,
metric = "ROC",
trControl = ctrl)
pred <- predict.train(rfFit, newdata = test, type = "prob")
I get the error: Error in [.data.frame(out, , obsLevels, drop = FALSE) :
undefined columns selected
The variable names on the test data set are:
str(test)
'data.frame': 3171 obs. of 29 variables:
$ alchemy_category : Factor w/ 13 levels "arts_entertainment",..: 8 4 12 4 10 12 12 8 1 2 ...
$ alchemy_category_score : num 5307 4825 1 6708 5416 ...
$ avglinksize : num 2.56 3.77 2.27 2.52 1.85 ...
$ commonlinkratio_1 : num 0.39 0.462 0.496 0.706 0.471 ...
$ commonlinkratio_2 : num 0.257 0.205 0.385 0.346 0.161 ...
$ commonlinkratio_3 : num 0.0441 0.0513 0.1709 0.123 0.0323 ...
$ commonlinkratio_4 : num 0.0221 0 0.1709 0.0906 0 ...
$ compression_ratio : num 0.49 0.782 1.25 0.449 0.454 ...
$ embed_ratio : num 0 0 0 0 0 0 0 0 0 0 ...
$ frameTagRatio : num 0.0671 0.0429 0.0588 0.0581 0.093 ...
$ hasDomainLink : Factor w/ 2 levels "0","1": 1 1 1 1 1 1 1 1 1 1 ...
$ html_ratio : num 0.23 0.366 0.162 0.147 0.244 ...
$ image_ratio : num 0.19944 0.08 10 0.00596 0.03571 ...
$ is_news : Factor w/ 2 levels "0","1": 2 1 1 2 2 1 1 2 1 1 ...
$ lengthyLinkDomain : Factor w/ 2 levels "0","1": 2 2 2 2 1 2 2 1 1 1 ...
$ linkwordscore : num 15 62 42 41 34 35 15 22 41 7 ...
$ news_front_page : Factor w/ 2 levels "0","1": 1 1 1 1 1 1 1 1 1 1 ...
$ non_markup_alphanum_characters: num 5643 382 2420 5559 2209 ...
$ numberOfLinks : num 136 39 117 309 155 266 55 145 110 1 ...
$ numwords_in_url : num 3 2 1 10 10 7 1 9 5 0 ...
$ parametrizedLinkRatio : num 0.2426 0.1282 0.5812 0.0388 0.0968 ...
$ spelling_errors_ratio : num 0.0806 0.1765 0.125 0.0631 0.0653 ...
$ isVideo : Factor w/ 2 levels "0","1": 1 2 1 2 2 2 1 1 2 2 ...
$ isFashion : Factor w/ 2 levels "0","1": 1 1 1 1 1 2 1 1 1 1 ...
$ isFood : Factor w/ 2 levels "0","1": 2 2 2 2 2 2 2 2 2 2 ...
$ hasComments : Factor w/ 2 levels "0","1": 2 1 1 2 2 2 1 2 2 1 ...
$ hasGoogleAnalytics : Factor w/ 2 levels "0","1": 1 2 2 2 2 1 1 2 1 1 ...
$ hasInlineCSS : Factor w/ 2 levels "0","1": 2 2 2 1 1 2 2 2 1 1 ...
$ noOfMetaTags : num 3 6 5 9 16 22 6 9 7 0 ...
If I omit the type="prob" part, I get no error.
Any ideas?
Could it be the length of the variable "alchemy_category" which is appended with the respective factor levels e.g. "alchemy_categoryarts_entertainment" inside the model??

Answer 1

The answer is in bold at the top of your post =]
What are you modeling? Is it alchemy_category? The code only says formula and we can't see it.
When you ask for class probabilities, model predictions are a data frame with separate columns for each class/level. If alchemy_category doesn't have levels that are valid column names, data.frame converts then to valid names. That creates a problem because the code is looking for a specific name but the data frame as a different (but valid) name.
For example, if I had
> test <- factor(c("level1", "level 2"))
> levels(test)
[1] "level 2" "level1"
> make.names(levels(test))
[1] "level.2" "level1"
the code would be looking for "level 2" but there is only "level.2".

Answer 2

As stated above the class values must be factors and must be valid names. Another way to insure this is,
levels(all.dat$target) <- make.names(levels(factor(all.dat$target)))

Answer 3

I have read through the answers above while facing a similar problem. A formal solution is to do this on the train and test datasets. Make sure you include the response variable in the feature.names too.
feature.names=names(train)
for (f in feature.names) {
if (class(train[[f]])=="factor") {
levels <- unique(c(train[[f]]))
train[[f]] <- factor(train[[f]],
labels=make.names(levels))
}
}
This creates syntactically correct labels for all factors.

Answer 4

As #Sam Firke already pointed out in comments (but I overlooked it) levels TRUE/FALSE also don't work. So I converted them to yes/no.

Answer 5

As per the above example, usually refactoring the outcome variable will fix the problem. It's better to change in the original dataset before partitioning into training and test datasets
levels <- unique(data$outcome)
data$outcome <- factor(data$outcome, labels=make.names(levels))
As others pointed out earlier, this problem only occurs when classProbs=TRUE which causes the train function to generate additional statistics related to the outcome class