Generate all possible pairs and count frequency in R

Question

I have a data frame of products (apple, pear, banana) sold across different locations (cities) within different categories (food and edibles).

I would like to count how many times any given pair of products appeared together in any category.

This is an example dataset I'm trying to make this to work on:

category <- c('food','food','food','food','food','food','edibles','edibles','edibles','edibles', 'edibles')
location <- c('houston, TX', 'houston, TX', 'las vegas, NV', 'las vegas, NV', 'philadelphia, PA', 'philadelphia, PA', 'austin, TX', 'austin, TX', 'charlotte, NC', 'charlotte, NC', 'charlotte, NC')
item <- c('apple', 'banana', 'apple', 'pear', 'apple', 'pear', 'pear', 'apple', 'apple', 'pear', 'banana')

food_data <- data.frame(cbind(category, location, item), stringsAsFactors = FALSE)

For example, the pair "apple & banana" appeared together in the "food" category in "las vegas, NV", but also in the "edibles" category in "charlotte, NC". Therefore, the count for the "apple & banana" pair would be 2.

My desired output is count of pairs like this:

(unordered) count of apple & banana

2

(unordered) count of apple & pear

4

Anyone have an idea for how to accomplish this? Relatively new to R and have been confused for a while.

I'm trying to use this to calculate affinities between different items.

Additional clarification on output: My full dataset consists of hundreds of different items. Would like to get a data frame where the first column is the pair and the second column is the count for each pair.

Answer 1

Here is one way using tidyverse and crossprod ; By using spread , it turns all item/fruit from the same category-location combination into one row with the item as headers (this requires you have no duplicated item in each category-country, otherwise you need a pre-aggregation step), values indicating existence; crossprod essentially evaluates the inner product of pairs of items columns and gives the number of cooccurrences.

library(tidyverse)
food_data %>% 
    mutate(n = 1) %>% 
    spread(item, n, fill=0) %>% 
    select(-category, -location) %>% 
    {crossprod(as.matrix(.))} %>% 
    `diag<-`(0)

#       apple banana pear
#apple      0      2    4
#banana     2      0    1
#pear       4      1    0

---

To convert this to a data frame:

food_data %>% 
    mutate(n = 1) %>% 
    spread(item, n, fill=0) %>% 
    select(-category, -location) %>% 
    {crossprod(as.matrix(.))} %>% 
    replace(lower.tri(., diag=T), NA) %>%
    reshape2::melt(na.rm=T) %>%
    unite('Pair', c('Var1', 'Var2'), sep=", ")

#           Pair value
#4 apple, banana     2
#7   apple, pear     4
#8  banana, pear     1

Answer 2

A solution from the tidyverse . The idea is to create food_data2 , which is the wide format of food_data . After that, create the combination between each unique item and use map2_int to loop through each item combination to count the number. This solution should work for any numbers of items.

library(tidyverse)

food_data2 <- food_data %>%
  mutate(count = 1) %>%
  spread(item, count, fill = 0) 

food_combination <- food_data %>%
  pull(item) %>%
  unique() %>%
  combn(2) %>%
  t() %>%
  as_data_frame() %>%
  mutate(count = map2_int(V1, V2, 
                         ~sum(apply(food_data2 %>% select(.x, .y), 1, sum) == 2)))

# View the result
food_combination
# A tibble: 3 x 3
      V1     V2 count
   <chr>  <chr> <int>
1  apple banana     2
2  apple   pear     4
3 banana   pear     1

If you just want one column to show the item combination at the end, you can further use the unite function.

food_combination2 <- food_combination %>%
  unite(Pair, V1, V2)

# View the result
food_combination2
# A tibble: 3 x 2
          Pair count
*        <chr> <int>
1 apple_banana     2
2   apple_pear     4
3  banana_pear     1

Answer 3

Here is a little function that will do what you need. It could be generalized to arbitrary grouping columns with the dplyr:: evaluation system described here . Probably better ways to do it but this works :p

Comments/explanations are inline ~~

library("dplyr")

# a function to apply to `food_data` from the original post 
count_combos <- function(df, group_col1, group_col2, count_col){ 

  # use `combn()` to get all the unique pairs from the `$items` col
  combos <- t(combn(sort(unique(df[[count_col]])), 2)) %>% 
    as_data_frame() %>% 
    # initialize an empty column to catch the counts 
    mutate(count=NA)

  # create a new df from the colnames passed as args, 
  # (it would be more general to just use the dplyr evaluation system (@_@))
  df <- data_frame(
    group_col1 = df[[group_col1]],
    group_col2 = df[[group_col2]],
    count_col  = df[[count_col]]
  )
  # for each combo of the grouping vars, get a pipe-seperated string of items
  df <- df %>% 
    group_by(group_col1, group_col2) %>% summarize(
      items = paste(unique(count_col), collapse="|")
    ) %>% ungroup()

  # for each item pair/combo, get the number of rows of `df` with both items 
  combos$count <- sapply(1:nrow(combos), function(x){
    sum(grepl(combos$V1[x], df$items) & grepl(combos$V2[x], df$items))
  })
  # and return it in a nice df
  return(combos)
}

# apply the function 
count_combos(food_data, 
             group_col1="category", group_col2="location", count_col="item")