k-fold cross-validation in dplyr?

Question

Hadley Wickham proposed that it could be possible to conduct bootstrap using dplyr package, his suggestion was improved and then implemented in broom package. Would it be also possible to implement k-fold cross-validation?

I guess that the first step (selecting the train group) is very simple:

crossvalidate <- function (df, k = 5) {
  n <- nrow(df)
  idx <- sample(rep_len(1:k, n))
  attr(df, "indices") <- lapply(1:k, function(i) which(idx != i))
  attr(df, "drop") <- TRUE
  attr(df, "group_sizes") <- nrow(df) - unclass(table(idx))
  attr(df, "biggest_group_size") <- max(attr(df, "group_sizes"))
  attr(df, "labels") <- data.frame(replicate = 1:k)
  attr(df, "vars") <- list(quote(replicate))
  class(df) <- c("grouped_df", "tbl_df", "tbl", "data.frame")
  df
}

But somehow I cannot find anywhere any documentation of attr(, "indices") to learn if it is somehow possible to use indices "other that" the ones selected for selecting the test group indices. Do you have any ideas?

Answer 1

https://rpubs.com/dgrtwo/cv-modelr - there you have an example of k-fold cross-validation using dplyr package:

library(ISLR)
library(dplyr)
library(purrr)
library(modelr)
library(broom)
library(tidyr)

set.seed(1)

models <- Smarket %>%
  select(Today, Lag1:Lag5) %>%
  crossv_kfold(k = 20) %>%
  mutate(model = map(train, ~ lm(Today ~ ., data = .)))

predictions <- models %>%
  unnest(map2(model, test, ~ augment(.x, newdata = .y)))

predictions %>%
  summarize(MSE = mean((Today - .fitted) ^ 2),
            MSEIntercept = mean((Today - mean(Today))^2))

Answer 2

Here is one solution for stratified 5-fold CV using dplyr:

df_fold = df %>%
  group_by(group_var) %>%
  sample_frac(1) %>%
  mutate(fold=rep(1:5, length.out=n())) %>%
  ungroup

for(i in 1:5){
  val = df_fold %>% filter(fold==i)
  tr = df_fold %>% anti_join(val, by=ID_var)
}