For each row return the column name of the largest value

Question

I have a roster of employees, and I need to know at what department they are in most often. It is trivial to tabulate employee ID against department name, but it is trickier to return the department name, rather than the number of roster counts, from the frequency table. A simple example below (column names = departments, row names = employee ids).

DF <- matrix(sample(1:9,9),ncol=3,nrow=3)
DF <- as.data.frame.matrix(DF)
> DF
  V1 V2 V3
1  2  7  9
2  8  3  6
3  1  5  4

Now how do I get

> DF2
  RE
1 V3
2 V1
3 V2

Answer 1

One option using your data (for future reference, use set.seed() to make examples using sample reproducible):

DF <- data.frame(V1=c(2,8,1),V2=c(7,3,5),V3=c(9,6,4))

colnames(DF)[apply(DF,1,which.max)]
[1] "V3" "V1" "V2"

A faster solution than using apply might be max.col:

colnames(DF)[max.col(DF,ties.method="first")]
#[1] "V3" "V1" "V2"

...where ties.method can be any of "random" "first" or "last"

This of course causes issues if you happen to have two columns which are equal to the maximum. I'm not sure what you want to do in that instance as you will have more than one result for some rows. E.g.:

DF <- data.frame(V1=c(2,8,1),V2=c(7,3,5),V3=c(7,6,4))
apply(DF,1,function(x) which(x==max(x)))

[[1]]
V2 V3 
 2  3 

[[2]]
V1 
 1 

[[3]]
V2 
 2 

Answer 2

One solution could be to reshape the date from wide to long putting all the departments in one column and counts in another, group by the employer id (in this case, the row number), and then filter to the department(s) with the max value. There are a couple of options for handling ties with this approach too.

library(tidyverse)

# sample data frame with a tie
df <- data_frame(V1=c(2,8,1),V2=c(7,3,5),V3=c(9,6,5))

# If you aren't worried about ties:  
df %>% 
  rownames_to_column('id') %>%  # creates an ID number
  gather(dept, cnt, V1:V3) %>% 
  group_by(id) %>% 
  slice(which.max(cnt)) 

# A tibble: 3 x 3
# Groups:   id [3]
  id    dept    cnt
  <chr> <chr> <dbl>
1 1     V3       9.
2 2     V1       8.
3 3     V2       5.


# If you're worried about keeping ties:
df %>% 
  rownames_to_column('id') %>%
  gather(dept, cnt, V1:V3) %>% 
  group_by(id) %>% 
  filter(cnt == max(cnt)) %>% # top_n(cnt, n = 1) also works
  arrange(id)

# A tibble: 4 x 3
# Groups:   id [3]
  id    dept    cnt
  <chr> <chr> <dbl>
1 1     V3       9.
2 2     V1       8.
3 3     V2       5.
4 3     V3       5.


# If you're worried about ties, but only want a certain department, you could use rank() and choose 'first' or 'last'
df %>% 
  rownames_to_column('id') %>%
  gather(dept, cnt, V1:V3) %>% 
  group_by(id) %>% 
  mutate(dept_rank  = rank(-cnt, ties.method = "first")) %>% # or 'last'
  filter(dept_rank == 1) %>% 
  select(-dept_rank) 

# A tibble: 3 x 3
# Groups:   id [3]
  id    dept    cnt
  <chr> <chr> <dbl>
1 2     V1       8.
2 3     V2       5.
3 1     V3       9.

# if you wanted to keep the original wide data frame
df %>% 
  rownames_to_column('id') %>%
  left_join(
    df %>% 
      rownames_to_column('id') %>%
      gather(max_dept, max_cnt, V1:V3) %>% 
      group_by(id) %>% 
      slice(which.max(max_cnt)), 
    by = 'id'
  )

# A tibble: 3 x 6
  id       V1    V2    V3 max_dept max_cnt
  <chr> <dbl> <dbl> <dbl> <chr>      <dbl>
1 1        2.    7.    9. V3            9.
2 2        8.    3.    6. V1            8.
3 3        1.    5.    5. V2            5.

Answer 3

Based on the above suggestions, the following data.table solution worked very fast for me:

library(data.table)

set.seed(45)
DT <- data.table(matrix(sample(10, 10^7, TRUE), ncol=10))

system.time(
  DT[, col_max := colnames(.SD)[max.col(.SD, ties.method = "first")]]
)
#>    user  system elapsed 
#>    0.15    0.06    0.21
DT[]
#>          V1 V2 V3 V4 V5 V6 V7 V8 V9 V10 col_max
#>       1:  7  4  1  2  3  7  6  6  6   1      V1
#>       2:  4  6  9 10  6  2  7  7  1   3      V4
#>       3:  3  4  9  8  9  9  8  8  6   7      V3
#>       4:  4  8  8  9  7  5  9  2  7   1      V4
#>       5:  4  3  9 10  2  7  9  6  6   9      V4
#>      ---                                       
#>  999996:  4  6 10  5  4  7  3  8  2   8      V3
#>  999997:  8  7  6  6  3 10  2  3 10   1      V6
#>  999998:  2  3  2  7  4  7  5  2  7   3      V4
#>  999999:  8 10  3  2  3  4  5  1  1   4      V2
#> 1000000: 10  4  2  6  6  2  8  4  7   4      V1

And also comes with the advantage that can always specify what columns .SD should consider by mentioning them in .SDcols:

DT[, MAX2 := colnames(.SD)[max.col(.SD, ties.method="first")], .SDcols = c("V9", "V10")]

---

In case we need the column name of the smallest value, as suggested by @lwshang, one just needs to use -.SD:

DT[, col_min := colnames(.SD)[max.col(-.SD, ties.method = "first")]]

Answer 4

If you're interested in a data.table solution, here's one. It's a bit tricky since you prefer to get the id for the first maximum. It's much easier if you'd rather want the last maximum. Nevertheless, it's not that complicated and it's fast!

Here I've generated data of your dimensions (26746 * 18).

Data

set.seed(45)
DF <- data.frame(matrix(sample(10, 26746*18, TRUE), ncol=18))

data.table answer:

require(data.table)
DT <- data.table(value=unlist(DF, use.names=FALSE), 
            colid = 1:nrow(DF), rowid = rep(names(DF), each=nrow(DF)))
setkey(DT, colid, value)
t1 <- DT[J(unique(colid), DT[J(unique(colid)), value, mult="last"]), rowid, mult="first"]

Benchmarking:

# data.table solution
system.time({
DT <- data.table(value=unlist(DF, use.names=FALSE), 
            colid = 1:nrow(DF), rowid = rep(names(DF), each=nrow(DF)))
setkey(DT, colid, value)
t1 <- DT[J(unique(colid), DT[J(unique(colid)), value, mult="last"]), rowid, mult="first"]
})
#   user  system elapsed 
#  0.174   0.029   0.227 

# apply solution from @thelatemail
system.time(t2 <- colnames(DF)[apply(DF,1,which.max)])
#   user  system elapsed 
#  2.322   0.036   2.602 

identical(t1, t2)
# [1] TRUE

It's about 11 times faster on data of these dimensions, and data.table scales pretty well too.

---

Edit: if any of the max ids is okay, then:

DT <- data.table(value=unlist(DF, use.names=FALSE), 
            colid = 1:nrow(DF), rowid = rep(names(DF), each=nrow(DF)))
setkey(DT, colid, value)
t1 <- DT[J(unique(colid)), rowid, mult="last"]

Answer 5

One option from dplyr 1.0.0 could be:

DF %>%
 rowwise() %>%
 mutate(row_max = names(.)[which.max(c_across(everything()))])

     V1    V2    V3 row_max
  <dbl> <dbl> <dbl> <chr>  
1     2     7     9 V3     
2     8     3     6 V1     
3     1     5     4 V2     

In some contexts, it could be safer to use pmap() (requires purrr):

DF %>%
 mutate(row_max = pmap_chr(across(everything()), ~ names(c(...)[which.max(c(...))])))

Sample data:

DF <- structure(list(V1 = c(2, 8, 1), V2 = c(7, 3, 5), V3 = c(9, 6, 
4)), class = "data.frame", row.names = c(NA, -3L))

Answer 6

A dplyr solution:

Idea:

• add rowids as a column
• reshape to long format
• filter for max in each group

Code:

DF = data.frame(V1=c(2,8,1),V2=c(7,3,5),V3=c(9,6,4))
DF %>% 
  rownames_to_column() %>%
  gather(column, value, -rowname) %>%
  group_by(rowname) %>% 
  filter(rank(-value) == 1) 

Result:

# A tibble: 3 x 3
# Groups:   rowname [3]
  rowname column value
  <chr>   <chr>  <dbl>
1 2       V1         8
2 3       V2         5
3 1       V3         9

This approach can be easily extended to get the top n columns. Example for n=2:

DF %>% 
  rownames_to_column() %>%
  gather(column, value, -rowname) %>%
  group_by(rowname) %>% 
  mutate(rk = rank(-value)) %>%
  filter(rk <= 2) %>% 
  arrange(rowname, rk) 

Result:

# A tibble: 6 x 4
# Groups:   rowname [3]
  rowname column value    rk
  <chr>   <chr>  <dbl> <dbl>
1 1       V3         9     1
2 1       V2         7     2
3 2       V1         8     1
4 2       V3         6     2
5 3       V2         5     1
6 3       V3         4     2

Answer 7

This is a fast and simple tidyverse solution, that can easily be applied to any subset of columns in a data.frame. The version below also uses ifelse to add missing values if all columns are 0. The missing values would be useful if, e.g., someone wants to use it to recombine one-hot encoded columns. It works on the data in the question, but here's an example of a one-hot encoded data set that it also works on.

data <- data.frame(
   oh_a = c(1,0,0,1,0,0)
  ,oh_b = c(0,1,1,0,0,0)
  ,oh_c = c(0,0,0,0,1,0)
  ,d = c("l","m","n","o","p","q"))

f <- function(x){ifelse(rowSums(x)==0, NA, names(x)[max.col(x, "first")])}
data %>% 
  mutate(transformed = f(across(starts_with("oh"))))

output:

  oh_a oh_b oh_c d transformed
1    1    0    0 l        oh_a
2    0    1    0 m        oh_b
3    0    1    0 n        oh_b
4    1    0    0 o        oh_a
5    0    0    1 p        oh_c
6    0    0    0 q        <NA>

Answer 8

A simple for loop can also be handy:

> df<-data.frame(V1=c(2,8,1),V2=c(7,3,5),V3=c(9,6,4))
> df
  V1 V2 V3
1  2  7  9
2  8  3  6
3  1  5  4
> df2<-data.frame()
> for (i in 1:nrow(df)){
+   df2[i,1]<-colnames(df[which.max(df[i,])])
+ }
> df2
  V1
1 V3
2 V1
3 V2

Answer 9

Here is an answer that works with data.table and is simpler. This assumes your data.table is named yourDF:

j1 <- max.col(yourDF[, .(V1, V2, V3, V4)], "first")
yourDF$newCol <- c("V1", "V2", "V3", "V4")[j1]

Replace ("V1", "V2", "V3", "V4") and (V1, V2, V3, V4) with your column names

Answer 10

This one is fast:

with(DF, {
  names(DF)[(V1 > V2 & V1 > V3) * 1 + (V2 > V3 & V2 > V1) * 2 + (V3 > V1 & V3 > V2)*3]
})