R：ADA：如何在具有分类描述符的数据帧上使用对？

Question

I'm using the ada package for classification where the descriptor variables are both categoric and numeric . This leads to a problem when calling the pairs function. Below is an example which illustrates my problem:

require(rpart)
require(ada)

data(car90, package = "rpart")
target = "Country"
input = setdiff(names(car90), target)

# ada only works with two distinct responses
car90 = car90[car90$Country %in% c("USA", "Japan/USA"), ] 

# remove surplus factor levels
car90$Country = as.character(car90$Country) 

adaCar90 = ada(car90[, input], car90[, target])
pairs(adaCar90, car90[, input], vars = 32:33)

# Error in pairs.default(as.matrix(rbind(train.data, test.x))[, vars],   
# lower.panel = panel.low,  :  non-numeric argument to 'pairs'  

Selecting only numeric descriptors using the vars argument of the pairs function doesn't seem to resolve matters. Does anyone know how I can fix this?

Cheers.

Answer 1

It looks like the maintainer of the ada package didn't anticipate your specific use case. The very last line of ada:::pairs.ada reads:

pairs(as.matrix(rbind(train.data, test.x))[, vars], lower.panel = panel.low, 
    upper.panel = panel.up)

The problem lies in where [, vars] has been placed. The code binds together train.data and test.x , then turns the whole thing into a matrix, and then subsets. Because your train.data contains a bunch of non-numeric columns, as.matrix returns a character matrix. If you change the last line to this:

pairs(as.matrix(rbind(train.data, test.x)[, vars]), lower.panel = panel.low, 
    upper.panel = panel.up)

then as.matrix is only called on the subset that contains numeric data, and the function works.

EDIT

I think what I suggested above is a good long-term solution, but there could be an easier short-term fix: pass only those columns of your training data that you will need for the graph instead of using the vars option. That way, only numeric data is ever passed to that final line of code. That would probably get you your graphs without you needing to hack the function.