Bdtjtk

I am trying to compute the variable importance out of a caret 'loclda' model.

My training dataset contains 20 numeric predictors and 4 discrete outcome classes (2500 samples).

I built the model and computed the variable importance with the following commands:

# fit the model

control <- trainControl(

                    method="repeatedcv", 

                    number=10, 

                    repeats=3,

                    savePredictions='final',

                    summaryFunction=multiClassSummary,

                    returnData=F,

                    trim=T,

                    allowParallel=F

                )



my_loclda_model <- train(MyClass ~ ., data=trainData, method='loclda', trControl=control, importance=T)



# compute variable importance

varImp(my_loclda_model)

Error : is.list(x) is not TRUE



traceback()

7: stop(msg, call. = FALSE, domain = NA)

6: stopifnot(is.list(x), is.list(val))

5: modifyList(data, lapply(data[, fc], as.numeric))

4: asNumeric(x)

3: filterVarImp(x_dat, y_dat, nonpara = nonpara, ...)

2: varImp.train(model_name)

1: varImp(model_name)

From caret user manual we can read the following:

For models that do not have corresponding varImp methods, see filterVarImp

Assuming loclda model does not have a varImp method, I tried to use the filterVarImp function directly from the final model:

str(my_loclda_model$finalModel)

List of 12

 $ learn           : num [1:2500, 1:20] -0.404 -0.336 -0.655 -0.618 -0.151 ...

  ..- attr(*, "dimnames")=List of 2

  .. ..$ : chr [1:2500] "X2" "X3" "X5" "X6" ...

  .. ..$ : chr [1:20] "indexA1" "indexF1" "index4" "index5" ...

 $ grouping        : Factor w/ 4 levels "group1","group2","group3",..: 1 5 2 5 1 1 3 4 3 1 ...

  ..- attr(*, "names")= chr [1:2500] "2" "3" "5" "6" ...

 $ lev             : chr [1:4] "group1" "group2" "group3" ...

 $ weight.func     :function (x)

 $ k               : num 3401

 $ weighted.apriori: logi TRUE

 $ call            : language loclda(x = x, grouping = y, k = floor(param$k), importance = ..1)

 $ xNames          : chr [1:20] "indexA1" "indexF1" "index4" "index5" ...

 $ problemType     : chr "Classification"

 $ tuneValue       :'data.frame':   1 obs. of  1 variable:

  ..$ k: num 3401

 $ obsLevels       : atomic [1:4] group1 group2 group3 ...

  ..- attr(*, "ordered")= logi FALSE

 $ param           :List of 1

  ..$ importance: logi TRUE

 - attr(*, "class")= chr "loclda"





imp_out <- filterVarImp(x=my_loclda_model$finalModel$learn, y=my_loclda_model$finalModel$grouping)



Error in data[, fc] : (subscript) logical subscript too long

traceback()

5: lapply(data[, fc], as.numeric)

4: stopifnot(is.list(x), is.list(val))

3: modifyList(data, lapply(data[, fc], as.numeric))

2: asNumeric(x)

1: filterVarImp(x = model_name$finalModel$learn, y = model_name$finalModel$grouping)

I don't understand. What x and y arguments should I pass to filterVarImp?

Thanks !

Note: the same problem also happened when using other methods (e.g. svmRadial, svmPoly, LogitBoost, regLogistic, ...).

Specifying x=as.data.frame(my_loclda_model$finalModel$learn) instead of a matrix seems to solve the problem, although the manual mentions that both matrix and data frame can be used.