CODEWITHSUNDEEP
rplotdecision-tree
user989762
user989762

Reputation: 1694

in R: Error in is.data.frame(data) : object '' not found, C5.0 plot

This question is similar to some other questions on Stackoverflow (here, here and here), but different enough so that I cannot extrapolate those answers to my case.

I have a function in which I fit a C5.0 model and than try to plot the model.

train_d <- globald[train_ind,c(features,21)]
model <- C5.0(binclass ~ .,data=train_d,trials=10)

binclass is a column name in my training/test data (globald is a dataframe from which I subset rows with _ind indices and columns c(3:12,21), where column 21 is named binclass). Fitting works well. However, when I also add the line

plot(model,trial=0)

then I get the following error: Error in is.data.frame(data) : object 'train_d' not found.

How is it possible that when fitting the model, train_d is found and used correctly, but while plotting, train_d is nowhere to be found? And, any suggestion of how to solve this issue. Namespaces in [r] remain a mystery to me.

A minimal running example is the following:

f <- function(){
    library(C50)
    set.seed(1)
    class = c(1,2)
    d <- data.frame(feature1 = sample(1:10,10,replace=TRUE), feature2 = 1:10, binclass = class)
    d$binclass <- as.factor(d$binclass)
    model <- C5.0(binclass ~ ., data=d)
    plot(model)   
}

Calling f() results in the following error: Error in is.data.frame(data) : object 'd' not found

Edit: As per the answer from MrFlick, it seems that the cause of this problem is a bug in the C5.0 code. There are some workarounds are indicated by Pascal and MrFlick.

Upvotes: 7

Views: 54380

Answers (3)

user10854911
user10854911

Reputation: 31

You can use the special assignment operator <<- instead of the standard one (<-). It will save the object to the global environment and that can solve your problem.

Upvotes: 3

Cybernetic
Cybernetic

Reputation: 13334

@MrFlick almost had it but not quite. This problem for plotting is particularly annoying when trying to pass arbitrary data and target features to the C50 method. As pointed out by MrFlick it was to do with renaming terms. By renaming the x and y terms in the method call the plotting function won't get confused. 

tree_model$call$x <- data_train[, -target_index]
tree_model$call$y <- data_train[[target_feature]]

For example, here is a method for passing in arbitrary data and a target feature and still being able to plot the result: 

boosted_trees <- function(data_train, target_feature, iter_choice) {

    target_index <- grep(target_feature, colnames(data_train))
    model_boosted <- C5.0(x = data_train[, -target_index], y = data_train[[target_feature]], trial=iter_choice)
    model_boosted$call$x <- data_train[, -target_index]
    model_boosted$call$y <- data_train[[target_feature]]
    return(model_boosted)

}

The model object returned by the above method can be plotted as normal. 

model <- boosted_trees(data_train, 'my_target', 10)
plot(model)

Upvotes: 2

MrFlick
MrFlick

Reputation: 206232

There does appear to be a bug in the code when it comes to evaluating the command in the proper environment. The problem appears to be in the C50::model.frame.C5.0 function. The "cleanest" work around I could find was to add a terms property to your model. This will help encapsulate the function environment.

f <- function(){
    library(C50)
    set.seed(1)
    class = c(1,2)
    d <- data.frame(feature1 = sample(1:10,10,replace=TRUE), feature2 = 1:10, binclass = class)
    d$binclass <- as.factor(d$binclass)
    model <- C5.0(binclass ~ ., data=d)
    model$terms <- eval(model$call$formula)   #<---- Added line
    plot(model)   
}

Upvotes: 5

Related Questions