此R循环的更有效版本

Question

I'm used to Python and JS, and pretty new to R, but enjoying it for data analysis. I was looking to create a new field in my data frame, based on some if/else logic, and tried to do it in a standard/procedural way:

for (i in 1:nrow(df)) {
  if (is.na(df$First_Payment_date[i]) == TRUE) {
    df$User_status[i] = "User never paid"
  } else if (df$Payment_Date[i] >= df$First_Payment_date[i]) {
    df$User_status[i] = "Paying user"
  } else if (df$Payment_Date[i] < df$First_Payment_date[i]) {
    df$User_status[i] = "Attempt before first payment"
  } else {
    df$User_status[i] = "Error"
  }
}

But it was CRAZY slow. I tried running this on a data frame of ~3 million rows, and it took way, way too long. Any tips on the "R" way of doing this?

Note that the df$Payment_Date and df$First_Payment_date fields are formatted as dates.

Answer 1

If you initialize to "error" and then overwrite for the conditions enumerated using logical indexing this should be much faster. Those if(){}else{} statements for every row are killing you.

df$User_status <- "Error"
df$User_status[ is.na(df$First_Payment_date) ] <- "User never paid"
df$User_status[ df$Payment_Date >= df$First_Payment_date ] <- "Paying user"
df$User_status[ df$Payment_Date < df$First_Payment_date ] <- "Attempt before first payment"

Answer 2

I am benchmarking data.frame and data.table for relatively large dataset.

First we generate some data.

set.seed(1234)
library(data.table)
df = data.frame(First_Payment_date=c(sample(c(NA,1:100),1000000, replace=1)),
                 Payment_Date=c(sample(1:100,1000000, replace=1)))
dt = data.table(df)

Then set up the benchmark. I am testing between @BondedDust's answer and its data.table equivalence. I have slightly modified (debug) his code.

library(microbenchmark)

test_df = function(){
    df$User_status <- "Error"
    df$User_status[ is.na(df$First_Payment_date) ] <- "User never paid"
    df$User_status[ df$Payment_Date >= df$First_Payment_date ] <- "Paying user"
    df$User_status[ df$Payment_Date < df$First_Payment_date ] <- "Attempt before first payment"
}

test_dt = function(){
    dt[, User_status := "Error"]
    dt[is.na(First_Payment_date), User_status := "User never paid"]
    dt[Payment_Date >= First_Payment_date, User_status := "Paying user"]
    dt[Payment_Date < First_Payment_date, User_status := "Attempt before first payment"]
}

microbenchmark(test_df(), test_dt(), times=10)

The result: data.table is 4x faster than data.frame for the generated 1 million rows data.

> microbenchmark(test_df(), test_dt(), times=10)
Unit: milliseconds
      expr       min        lq    median        uq       max neval
 test_df() 247.29182 256.69067 287.89768 319.34873 330.33915    10
 test_dt()  66.74265  69.42574  70.27826  72.93969  80.89847    10

Note

data.frame is faster than data.table for small dataset (say, 10000 rows.)

Answer 3

I'm not certain that this will speed it up a lot, but you should see some improvement over the for loop you had before. The else 's aren't really necessary under these conditions.

Also, R has functions that act as for loops, and other types of loops. See ?apply .

Give this a shot, see how it works. I can't test it since we don't have your data.

> df$User_status[i] <- rep("Error", nrow(df)) 
      ## allocate a vector, fill it with "Error"

> sapply(seq(nrow(df)), function(i){

    if(is.na(df$First_Payment_date[i])){ 
      gsub("Error", "User never paid", df$User_status[i]) }

    if(df$Payment_Date[i] >= df$First_Payment_date[i]){
      gsub("Error", "Paying user", df$User_status[i]) }

    if (df$Payment_Date[i] < df$First_Payment_date[i]) {
      gsub("Error", "Attempt before first payment", df$User_status[i]) }

    })

Answer 4

The usual way to handle this sort of thing is via ifelse .

df$User_status <- with(df,
    ifelse(is.na(First_Payment_date), "User never paid",
    ifelse(Payment_Date >= First_Payment_date, "Paying user",
    ifelse(Payment_Date < First_Payment_date, "Attempt before first payment",
    "Error"))))