Skip to content

sample equal number of cases per class

2 messages · ollestrat, Rui Barradas

ollestrat
# 
Dear community

I have a dataframe and want to split it into a learn and a test partition.
However the learnset should be balanced, i.e. each class should have the
same number of cases. I tried and searched a lot, without success so far.
Maybe you can help?

Some example code
*# generate example data
df <- data.frame(class = as.factor(sample(1:3, 20, replace = T)), var1 =
rnorm(20,3), var2 = rnorm(20,6))
summary(df)

# split into learn and test sets using the caret package
require(caret)
ind <- createDataPartition(df$class, p=.8, list = F, times = 1)

# The problem is here: class sizes are not equal)
learnset <- df[ind,]
summary(learnset)*

Version info:
/> R.Version()
$platform
[1] "x86_64-pc-mingw32"
$arch
[1] "x86_64"
$os
[1] "mingw32"
$system
[1] "x86_64, mingw32"
$major
[1] "2"
$minor
[1] "15.1"/



--
View this message in context: http://r.789695.n4.nabble.com/sample-equal-number-of-cases-per-class-tp4648381.html
Sent from the R help mailing list archive at Nabble.com.
Rui Barradas
# 
Hello,

Function caret::createDatapartition preserves the proportions of 
classes, like its documentation says, so you should expected the result 
to be balanced only if the original data.frame is also balanced. A 
solution is to write a small function that chooses a balanced set of 
indices. Note that ths function below does _not_ use the same arguments 
as caret::createDataPartition, its arguments are:

x - the original vector, matrix or data.frame.
y - a vector, what to balance.
p - proportion of x to choose.


createSets <- function(x, y, p){
     nr <- NROW(x)
     size <- (nr * p) %/% length(unique(y))
     idx <- lapply(split(seq_len(nr), y), function(.x) sample(.x, size))
     unlist(idx)
}
ind <- createSets(df, df$class, 0.8)
lrn <- df[ind,]
summary(lrn)


Also, 'df' is a bad name for a variable, it allready is an R function. 
Use, for instance, 'dat'.

Hope this helps,

Rui Barradas
Em 04-11-2012 10:47, ollestrat escreveu: