Clear and Concise Way to apply Standardization to both Train and Test Set in R

Question

I am selecting a 90/10 Training/Test split with some data in R. After I have the Training set. I would like to standardize it. I would then like to use the same mean and standard deviation used in the training set and apply that standardization to the test set.

I would like to do this in the most base-R way possible but would be ok with a dplyr solution too. Note that I have columns that are both factors/chr and numeric. Of course I need to select the numeric ones first.

My first setup is below with a reproducible example code. I have the means and standard deviations for the appropriate numeric columns, now how can I apply the standardization back to the specific columns on the training and test data?

library(tidyverse)
rm(list = ls())
x <- data.frame("hame" =  c("Bob", "Roberta", "Brady", "Jen", "Omar", "Phillip", "Natalie", "Aaron", "Annie", "Jeff"),
                "age" = c(60, 55, 25, 30, 35, 40, 47, 32, 34,67),
                "income" = c(50000, 60000, 100000, 90000, 100000, 95000, 75000, 85000, 95000, 105000))

train_split_pct = 0.90

train_size <- ceiling(nrow(x)*train_split_pct)  # num of rows for training set
test_size <- nrow(x) - train_size               # num of rows for testing set 

set.seed(123)
ix <-  sample(1:nrow(x)) # shuffle
x_new = x[ix, ]
Train_set  = x_new[1:train_size, ]
Test_set   = x_new[(train_size+1):(train_size+test_size), ]

Train_mask <- Train_set %>% select_if(is.numeric) 
Train_means <- Train_mask %>% apply(2, mean)
Train_stddevs <- Train_mask %>% apply(2, sd)

Answer 1

So after reviewing the prior answers which worked fine, I found them a bit unclear to use and not intuitive. I have achieved the desired result via a for loop. While slightly rudimentary I believe it a more clear approach. Given the use case where I don't have many columns I don't see a major issue in this solution unless there were many columns of data to go through. In that case I would need help seeking a faster solution.

Regardless, my method is as follows. I gather all column names in my Train_mask which is only the numeric columns. Next, I loop through each of the names and update the values accordingly with the standardization from their respective Train_means and Train_stddevs.

Due to the way I construct my Training and Testing sets there should be no issues with the order of my column frames and they can be used sequentially in the following fashion.

library(tidyverse)
rm(list = ls())
x <- data.frame("name" =  c("Bob", "Roberta", "Brady", "Jen", "Omar", "Phillip", "Natalie", "Aaron", "Annie", "Jeff"),
                "age" = c(60, 55, 25, 30, 35, 40, 47, 32, 34,67),
                "income" = c(50000, 60000, 100000, 90000, 100000, 95000, 75000, 85000, 95000, 105000))

train_split_pct = 0.90

train_size <- ceiling(nrow(x)*train_split_pct)  # num of rows for training set
test_size <- nrow(x) - train_size               # num of rows for testing set 

set.seed(123)
ix <-  sample(1:nrow(x)) # shuffle
x_new = x[ix, ]
Train_set  = x_new[1:train_size, ]
Test_set   = x_new[(train_size+1):(train_size+test_size), ]

Train_mask <- Train_set %>% select_if(is.numeric) 
Train_means <- data.frame(as.list(Train_mask %>% apply(2, mean)))
Train_stddevs <- data.frame(as.list(Train_mask %>% apply(2, sd)))


col_names <- names(Train_mask)
for (i in 1:ncol(Train_mask)){
  Train_set[,col_names[i]] <- (Train_set[,col_names[i]] - Train_means[,col_names[i]])/Train_stddevs[,col_names[i]]
  Test_set[,col_names[i]] <-  (Test_set[,col_names[i]] - Train_means[,col_names[i]])/Train_stddevs[,col_names[i]]
}

Train_set
Test_set

Output:

> Train_set
      name       age     income
3    Brady -3.180620  0.7745967
10    Jeff -2.972814  1.0327956
2  Roberta -3.032187 -1.2909944
8    Aaron -3.145986  0.0000000
6  Phillip -3.106404  0.5163978
9    Annie -3.136090  0.5163978
1      Bob -3.007448 -1.8073922
7  Natalie -3.071769 -0.5163978
5     Omar -3.131143  0.7745967
> Test_set
  name        age    income
4  Jen -0.9769502 0.2581989

Answer 2

We can do this in a concise way. Get the mean, sd of the 'Train' dataset ('mean_sd'). Note that with dplyr version >= 1.0, summarise can return more than one row. So, make use of that feature to create a two row dataset - first row => mean, second row => sd

library(dplyr) # >= 1.0.0    
library(purrr)
mean_sd <- Train_set %>%
    summarise(across(where(is.numeric),  ~ c(mean(., na.rm = TRUE), 
            sd(., na.rm = TRUE))))

Then, create a function ('f1') to do the standardization.

f1 <- function(x, y) (x -y[1])/y[2]

Loop over the list of 'Train', 'Test' dataset, use map2 to loop over the corresponding columns based on the 'mean_sd' dataset, apply the f1 and assign that output to the columns. Then, with list2env, we can update the same objects in the global environment

list2env(map(lst(Train_set, Test_set), ~  {
   .x[names(mean_sd)] <- map2(select(.x, names(mean_sd)), mean_sd, f1)
         .x}), .GlobalEnv)

-output

Train_set
#   hame        age     income
#3    Brady -1.3286522  0.7745967
#10    Jeff  1.6256451  1.0327956
#2  Roberta  0.7815601 -1.2909944
#8    Aaron -0.8362693  0.0000000
#6  Phillip -0.2735460  0.5163978
#9    Annie -0.6955885  0.5163978
#1      Bob  1.1332622 -1.8073922
#7  Natalie  0.2188368 -0.5163978
#5     Omar -0.6252481  0.7745967


Test_set
# hame        age    income
#4  Jen -0.9769502 0.2581989

Answer 3

Consider this as an option. You can use scale() function that allows you to normalize your variables. At the end you can find the code. Also, you can use mutate_if() in order to choose the numeric variables and avoid creating other dataframes. Here the code using dplyr where I have created two new dataframes with the required values:

library(tidyverse)
rm(list = ls())
x <- data.frame("hame" =  c("Bob", "Roberta", "Brady", "Jen", "Omar", "Phillip", "Natalie", "Aaron", "Annie", "Jeff"),
                "age" = c(60, 55, 25, 30, 35, 40, 47, 32, 34,67),
                "income" = c(50000, 60000, 100000, 90000, 100000, 95000, 75000, 85000, 95000, 105000))

train_split_pct = 0.90

train.size <- ceiling(nrow(x)*train_split_pct)  # num of rows for training set
test.size <- nrow(x) - train.size               # num of rows for testing set 

set.seed(123)
ix <-  sample(1:nrow(x)) # shuffle
x_new = x[ix, ]
Train.set  = x_new[1:train.size, ]
Test.set   = x_new[(train.size+1):(train.size+test.size), ]
#Normalize
Train.set2 <- Train.set %>%
  mutate_if(is.numeric, scale)
Test.set2 <- Test.set %>%
  mutate_if(is.numeric, scale)

Update: If the scale() is not working, you can try reshaping the data and joining with the computed values for mean and SD:

#Define indexes for numeric vars
index.train <- which(names(Train.set)%in% names(Train_means))
#Format means and sd to merge
Train2 <- Train.set %>% 
  mutate(id=row_number()) %>%
  pivot_longer(cols=index.train) %>%
  left_join(
    Train_means %>% t() %>%data.frame %>%
      pivot_longer(everything()) %>%
      rename(Mean=value) %>%
      left_join(Train_stddevs %>% t() %>%data.frame %>%
                  pivot_longer(everything()) %>%
                  rename(SD=value))
  ) %>%
  #Compute standard values
  mutate(SValue=(value-Mean)/SD) %>%
  select(-c(value,Mean,SD)) %>%
  pivot_wider(names_from = name,values_from=SValue) %>% select(-id)

Output:

# A tibble: 9 x 3
  hame       age income
  <fct>    <dbl>  <dbl>
1 Brady   -1.33   0.775
2 Jeff     1.63   1.03 
3 Roberta  0.782 -1.29 
4 Aaron   -0.836  0    
5 Phillip -0.274  0.516
6 Annie   -0.696  0.516
7 Bob      1.13  -1.81 
8 Natalie  0.219 -0.516
9 Omar    -0.625  0.775

And for the test set, the process is similar:

#Define indexes
index.test <- which(names(Test.set)%in% names(Train_means))
#Format means and sd 2
Test2 <- Test.set %>% 
  mutate(id=row_number()) %>%
  pivot_longer(cols=index.test) %>%
  left_join(
    Train_means %>% t() %>%data.frame %>%
      pivot_longer(everything()) %>%
      rename(Mean=value) %>%
      left_join(Train_stddevs %>% t() %>%data.frame %>%
                  pivot_longer(everything()) %>%
                  rename(SD=value))
  ) %>%
  #Compute standard values
  mutate(SValue=(value-Mean)/SD) %>%
  select(-c(value,Mean,SD)) %>%
  pivot_wider(names_from = name,values_from=SValue) %>% select(-id)

Output:

# A tibble: 1 x 3
  hame     age income
  <fct>  <dbl>  <dbl>
1 Jen   -0.977  0.258

The key is merging the values after reshaping. As evidence I will show the intermediate step for the final dataset. It looks like this:

# A tibble: 2 x 7
  hame     id name   value    Mean      SD SValue
  <fct> <int> <chr>  <dbl>   <dbl>   <dbl>  <dbl>
1 Jen       1 age       30    43.9    14.2 -0.977
2 Jen       1 income 90000 85000   19365.   0.258

In that way is easy to compute the standard values you want.