data.table: “group counter” for a specific combination of columns - r

Question

I would like to add a counter column in a data frame based on a set of identical rows. To do this, I used the package data.table. In my case, the comparison between rows need doing from the combination of columns "z" AND ("x" OR "y").
I tested:
DF[ , Index := .GRP, by = c("x","y","z") ]
but the result is the combination of "z" AND "x" AND "y".
How can I have the combination of "z" AND ("x" OR "y") ?
Here is a data example:
DF = data.frame(x=c("a","a","a","b","c","d","e","f","f"), y=c(1,3,2,8,8,4,4,6,0), z=c("M","M","M","F","F","M","M","F","F"))
DF <- data.table(DF)
I would like to have this output:
> DF
x y z Index
1: a 1 M 1
2: a 3 M 1
3: a 2 M 1
4: b 8 F 2
5: c 8 F 2
6: d 4 M 3
7: e 4 M 3
8: f 6 F 4
9: f 0 F 4

Answer 1

The new group starts if the value for z is changing or the values both for x and y are changing.
Try this example.
require(data.table)
DF <- data.table(x = c("a","a","a","b","c","d","e","f","f"),
y = c(1,3,2,8,8,4,4,6,0),
z=c("M","M","M","F","F","M","M","F","F"))
# The functions to compare if value is not equal with the previous value
is.not.eq.with.lag <- function(x) c(T, tail(x, -1) != head(x, -1))
DF[, x1 := is.not.eq.with.lag(x)]
DF[, y1 := is.not.eq.with.lag(y)]
DF[, z1 := is.not.eq.with.lag(z)]
DF
DF[, Index := cumsum(z1 | (x1 & y1))]
DF

Answer 2

I know a lot of people warn against a for loop in R, but in this instance I think it is a very direct way of approaching the problem. Plus, the result isn't growing in size so performance issues aren't a large issue. The for loop approach would be:
dt$grp <- rep(NA,nrow(dt))
for (i in 1:nrow(dt)){
if (i == 1){
dt$grp[i] = 1
}
else {
if(dt$z[i-1] == dt$z[i] & (dt$x[i-1] == dt$x[i] | dt$y[i-1] == dt$y[i])){
dt$grp[i] = dt$grp[i-1]
}else{
dt$grp[i] = dt$grp[i-1] + 1
}
}
}
Trying this on OPs original problem, the result is:
DF = data.frame(x=c("a","a","a","b","c","d","e","f","f"), y=c(1,3,2,8,8,4,4,6,0), z=c("M","M","M","F","F","M","M","F","F"))
dt <- data.table(DF)
dt$grp <- rep(NA,nrow(dt))
for (i in 1:nrow(dt)){
if (i == 1){
dt$grp[i] = 1
}
else {
if(dt$z[i-1] == dt$z[i] & (dt$x[i-1] == dt$x[i] | dt$y[i-1] == dt$y[i])){
dt$grp[i] = dt$grp[i-1]
}else{
dt$grp[i] = dt$grp[i-1] + 1
}
}
}
dt
x y z grp
1: a 1 M 1
2: a 3 M 1
3: a 2 M 1
4: b 8 F 2
5: c 8 F 2
6: d 4 M 3
7: e 4 M 3
8: f 6 F 4
9: f 0 F 4
Trying this on the data.table in #Frank's comment, gives the expected result as well:
dt<-data.table(x = c("b", "a", "a"), y = c(1, 1, 2), z = c("F", "F", "F"))
dt$grp <- rep(NA,nrow(dt))
for (i in 1:nrow(dt)){
if (i == 1){
dt$grp[i] = 1
}
else {
if(dt$z[i-1] == dt$z[i] & (dt$x[i-1] == dt$x[i] | dt$y[i-1] == dt$y[i])){
dt$grp[i] = dt$grp[i-1]
}else{
dt$grp[i] = dt$grp[i-1] + 1
}
}
}
dt
x y z grp
1: b 1 F 1
2: a 1 F 1
3: a 2 F 1

Answer 3

EDITED TO ADD: This solution is in some ways a more verbose version of the one advocated by djhurio above. I think it shows what is happening a bit more so I'll leave it.
I think this is a task easier to do if it is broken down a little bit. The below code creates TWO indices at first, one for changes in x (nested in z) and one for changes in y (nested in z). We then find the first row from each of these indices. Taking the cumulative sum of the case where both FIRST.x and FIRST.y is true should give your desired index.
library(data.table)
dt_example <- data.table(x = c("a","a","a","b","c","d","e","f","f"),
y = c(1,3,2,8,8,4,4,6,0),
z = c("M","M","M","F","F","M","M","F","F"))
dt_example[,Index_x := .GRP,by = c("z","x")]
dt_example[,Index_y := .GRP,by = c("z","y")]
dt_example[,FIRST.x := !duplicated(Index_x)]
dt_example[,FIRST.y := !duplicated(Index_y)]
dt_example[,Index := cumsum(FIRST.x & FIRST.y)]
dt_example
x y z Index_x Index_y FIRST.x FIRST.y Index
1: a 1 M 1 1 TRUE TRUE 1
2: a 3 M 1 2 FALSE TRUE 1
3: a 2 M 1 3 FALSE TRUE 1
4: b 8 F 2 4 TRUE TRUE 2
5: c 8 F 3 4 TRUE FALSE 2
6: d 4 M 4 5 TRUE TRUE 3
7: e 4 M 5 5 TRUE FALSE 3
8: f 6 F 6 6 TRUE TRUE 4
9: f 0 F 6 7 FALSE TRUE 4

Answer 4

This approach looks for changes in x & z | y & z. The extra columns are left in the data.table to show the calculations.
DF[, c("Ix", "Iy", "Iz", "dx", "dy", "min.change", "Index") :=
#Create index of values based on consecutive order
list(ix <- rleid(x), iy <- rleid(y), iz <- rleid(z),
#Determine if combinations of x+z OR y+z change
ix1 <- c(0, diff(rleid(ix+iz))),
iy1 <- c(0, diff(rleid(iy+iz))),
#Either combination is constant (no change)?
change <- pmin(ix1, iy1),
#New index based on change
cumsum(change) + 1
)]
x y z Ix Iy Iz dx dy min.change Index
1: a 1 M 1 1 1 0 0 0 1
2: a 3 M 1 2 1 0 1 0 1
3: a 2 M 1 3 1 0 1 0 1
4: b 8 F 2 4 2 1 1 1 2
5: c 8 F 3 4 2 1 0 0 2
6: d 4 M 4 5 3 1 1 1 3
7: e 4 M 5 5 3 1 0 0 3
8: f 6 F 6 6 4 1 1 1 4
9: f 0 F 6 7 4 0 1 0 4