Row Differences in Dataframe by Group
-
21-12-2019 - |
Question
My problem has to do with finding row differences in a data frame by group. I've tried to do this a few ways. Here's an example. The real data set is several million rows long.
set.seed(314)
df = data.frame("group_id"=rep(c(1,2,3),3),
"date"=sample(seq(as.Date("1970-01-01"),Sys.Date(),by=1),9,replace=F),
"logical_value"=sample(c(T,F),9,replace=T),
"integer"=sample(1:100,9,replace=T),
"float"=runif(9))
df = df[order(df$group_id,df$date),]
I ordered it by group_id and date so that the diff function can find the sequential differences, which results in time ordered differences of the logical, integer, and float variables. I could easily do some sort of apply(df,2,diff), but I need it by group_id. Hence, doing apply(df,2,diff) results in extra unneeded results.
df
group_id date logical_value integer float
1 1 1974-05-13 FALSE 4 0.03472876
4 1 1979-12-02 TRUE 45 0.24493995
7 1 1980-08-18 TRUE 2 0.46662253
5 2 1978-12-08 TRUE 56 0.60039164
2 2 1981-12-26 TRUE 34 0.20081799
8 2 1986-05-19 FALSE 60 0.43928929
6 3 1983-05-22 FALSE 25 0.01792820
9 3 1994-04-20 FALSE 34 0.10905326
3 3 2003-11-04 TRUE 63 0.58365922
So I thought I could break up my data frame into chunks by group_id, and pass each chunk into a user defined function:
create_differences = function(data_group){
apply(data_group, 2, diff)
}
But I get errors using the code:
diff_df = lapply(split(df,df$group_id),create_differences)
Error in r[i1] - r[-length(r):-(length(r) - lag + 1L)] : non-numeric argument to binary operator
by(df,df$group_id,create_differences)
Error in r[i1] - r[-length(r):-(length(r) - lag + 1L)] : non-numeric argument to binary operator
As a side note, the data is nice, no NAs, nulls, blanks, and every group_id has at least 2 rows associated with it.
Edit 1: User alexis_laz correctly pointed out that my function needs to be sapply(data_group, diff).
Using this edit, I get a list of data frames (one list entry per group).
Edit 2:
The expected output would be a combined data frame of differences. Ideally, I would like to keep the group_id, but if not, it's not a big deal. Here is what the sample output should be like:
diff_df
group_id date logical_value integer float
[1,] 1 2029 1 41 0.2102112
[2,] 1 260 0 -43 0.2216826
[1,] 2 1114 0 -22 -0.3995737
[2,] 2 1605 -1 26 0.2384713
[1,] 3 3986 0 9 0.09112507
[2,] 3 3485 1 29 0.47460596
La solution
I think regarding the fact that you have millions of rows you can move to the data.table
suitable for by group actions.
library(data.table)
DT <- as.data.table(df)
## this will order per group and per day
setkeyv(DT,c('group_id','date'))
## for all column apply diff
DT[,lapply(.SD,diff),group_id]
# group_id date logical_value integer float
# 1: 1 2029 days 1 41 0.21021119
# 2: 1 260 days 0 -43 0.22168257
# 3: 2 1114 days 0 -22 -0.39957366
# 4: 2 1604 days -1 26 0.23847130
# 5: 3 3987 days 0 9 0.09112507
# 6: 3 3485 days 1 29 0.47460596
Autres conseils
It certainly won't be as quick compared to data.table
but below is an only slightly ugly base solution using aggregate
:
result <- aggregate(. ~ group_id, data=df, FUN=diff)
result <- cbind(result[1],lapply(result[-1], as.vector))
result[order(result$group_id),]
# group_id date logical_value integer float
#1 1 2029 1 41 0.21021119
#4 1 260 0 -43 0.22168257
#2 2 1114 0 -22 -0.39957366
#5 2 1604 -1 26 0.23847130
#3 3 3987 0 9 0.09112507
#6 3 3485 1 29 0.47460596