Cross-correlation with multiple groups in one data.table - r

I'd like to calculate the cross-correlations between groups of time series within on data.table. I have a time series data in this format:
data = data.table( group = c(rep("a", 5),rep("b",5),rep("c",5)) , Y = rnorm(15) )
group Y
1: a 0.90855520
2: a -0.12463737
3: a -0.45754652
4: a 0.65789709
5: a 1.27632196
6: b 0.98483700
7: b -0.44282527
8: b -0.93169070
9: b -0.21878359
10: b -0.46713392
11: c -0.02199363
12: c -0.67125826
13: c 0.29263953
14: c -0.65064603
15: c -1.41143837
Each group has the same number of observations. What I am looking for is a way to obtain cross correlation between the groups:
group.1 group.2 correlation
a b 0.xxx
a c 0.xxx
b c 0.xxx
I am working on a script to subset each group and append the cross-correlations, but the data size is fairly large. Is there any efficient / zen way to do this?

Does this help?
data[,id:=rep(1:5,3)]
dtw = dcast.data.table(data, id ~ group, value.var="Y" )[, id := NULL]
cor(dtw)
See Correlation between groups in R data.table
Another way would be:
# data
set.seed(45L)
data = data.table( group = c(rep("a", 5),rep("b",5),rep("c",5)) , Y = rnorm(15) )
# method 2
setkey(data, "group")
data2 = data[J(c("b", "c", "a"))][, list(group2=group, Y2=Y)]
data[, c(names(data2)) := data2]
data[, cor(Y, Y2), by=list(group, group2)]
# group group2 V1
# 1: a b -0.2997090
# 2: b c 0.6427463
# 3: c a -0.6922734
And to generalize this "other" way to more than three groups...
data = data.table( group = c(rep("a", 5),rep("b",5),rep("c",5),rep("d",5)) ,
Y = rnorm(20) )
setkey(data, "group")
groups = unique(data$group)
ngroups = length(groups)
library(gtools)
pairs = combinations(ngroups,2,groups)
d1 = data[pairs[,1],,allow.cartesian=TRUE]
d2 = data[pairs[,2],,allow.cartesian=TRUE]
d1[,c("group2","Y2"):=d2]
d1[,cor(Y,Y2), by=list(group,group2)]
# group group2 V1
# 1: a b 0.10742799
# 2: a c 0.52823511
# 3: a d 0.04424170
# 4: b c 0.65407400
# 5: b d 0.32777779
# 6: c d -0.02425053

Related

Create n new columns by n distinct groups in data.table

I have the following data table, and would like to sum y twice grouping the first time by g1 and the second time by g2.
Usually I would just chain the calculations together, but I would like to be able to do the grouped sum n different times by n groups.
library(data.table)
DT <- data.table(
g1 = c("a", "b"),
g2 = c("a", "a"),
y = c(3,5)
)
new_cols <- paste0("sum_by_", c("g1", "g2"))
group_cols <- c("g1", "g2")
# Supplying cols to by like this groups by g1 AND g2, when in reality I want it to
# take g1 the first time and g2 the second time.
DT[, paste(new_cols) := lapply(rep(y, length(new_cols)), sum),
by = .(group_cols)][]
this gives me:
# g1 g2 y sum_by_g1 sum_by_g2
# 1: a a 3 3 3
# 2: b a 5 5 5
when I actually want:
# g1 g2 y sum_by_g1 sum_by_g2
# 1: a a 3 3 8
# 2: b a 5 5 8
Is there any slick data.table way to do this? Something like supplying .SD to by (this in itself doesn't seem to work)?
Edit: Changed y from c(1,1) to c(3,5)
Edit Rationale: Actual and desired outputs while y = c(1,1) gave the impression that I wanted to count the observations in each group, when I actually want to sum(y) for each group.
The grouping should be separate as a a and a b are regarded as unique elements thus, there is only a single observation per group
for(i in seq_along(group_cols)) DT[, (new_cols[i]) := sum(y), by = c(group_cols[i])]
-output
DT
g1 g2 y sum_by_g1 sum_by_g2
1: a a 3 3 8
2: b a 5 5 8
You can try Reduce like below
> Reduce(function(dt, g) dt[, paste0("sum_by_", g) := .N, g], list(DT, "g1", "g2"))[]
g1 g2 y sum_by_g1 sum_by_g2
1: a a 1 1 2
2: b a 1 1 2
or
> Reduce(function(dt, g) dt[, paste0("sum_by_", g) := .N, g], c("g1", "g2"),init = DT)[]
g1 g2 y sum_by_g1 sum_by_g2
1: a a 1 1 2
2: b a 1 1 2

Merging a sum by reference with data.table

Let's say I have two data.table, dt_a and dt_b defined as below.
library(data.table)
set.seed(20201111L)
dt_a <- data.table(
foo = c("a", "b", "c")
)
dt_b <- data.table(
bar = sample(c("a", "b", "c"), 10L, replace=TRUE),
value = runif(10L)
)
dt_b[]
## bar value
## 1: c 0.4904536
## 2: c 0.9067509
## 3: b 0.1831664
## 4: c 0.0203943
## 5: c 0.8707686
## 6: a 0.4224133
## 7: a 0.6025349
## 8: b 0.4916672
## 9: a 0.4566726
## 10: b 0.8841110
I want to left join dt_b on dt_a by reference, summing over the multiple match. A way to do so would be to first create a summary of dt_b (thus solving the multiple match issue) and merge if afterwards.
dt_b_summary <- dt_b[, .(value=sum(value)), bar]
dt_a[dt_b_summary, value_good:=value, on=c(foo="bar")]
dt_a[]
## foo value_good
## 1: a 1.481621
## 2: b 1.558945
## 3: c 2.288367
However, this will allow memory to the object dt_b_summary, which is inefficient.
I would like to have the same result by directly joining on dt_b and summing multiple match. I'm looking for something like below, but that won't work.
dt_a[dt_b, value_bad:=sum(value), on=c(foo="bar")]
dt_a[]
## foo value_good value_bad
## 1: a 1.481621 5.328933
## 2: b 1.558945 5.328933
## 3: c 2.288367 5.328933
Anyone knows if there is something possible?
We can use .EACHI with by
library(data.table)
dt_b[dt_a, .(value = sum(value)), on = .(bar = foo), by = .EACHI]
# bar value
#1: a 1.481621
#2: b 1.558945
#3: c 2.288367
If we want to update the original object 'dt_a'
dt_a[, value := dt_b[.SD, sum(value), on = .(bar = foo), by = .EACHI]$V1]
dt_a
# foo value
#1: a 1.481621
#2: b 1.558945
#3: c 2.288367
For multiple columns
dt_b$value1 <- dt_b$value
nm1 <- c('value', 'value1')
dt_a[, (nm1) := dt_b[.SD, lapply(.SD, sum),
on = .(bar = foo), by = .EACHI][, .SD, .SDcols = nm1]]

Ranking multiple columns by different orders using data table

Using my example below, how can I rank multiple columns using different orders, so for example rank y as descending and z as ascending?
require(data.table)
dt <- data.table(x = c(rep("a", 5), rep("b", 5)),
y = abs(rnorm(10)) * 10, z = abs(rnorm(10)) * 10)
cols <- c("y", "z")
dt[, paste0("rank_", cols) := lapply(.SD, function(x) frankv(x, ties.method = "min")), .SDcols = cols, by = .(x)]
data.table's frank() function has some useful features which aren't available in base R's rank() function (see ?frank). E.g., we can reverse the order of the ranking by prepending the variable with a minus sign:
library(data.table)
# create reproducible data
set.seed(1L)
dt <- data.table(x = c(rep("a", 5), rep("b", 5)),
y = abs(rnorm(10)) * 10, z = abs(rnorm(10)) * 10)
# rank y descending, z ascending
dt[, rank_y := frank(-y), x][, rank_z := frank(z), x][]
x y z rank_y rank_z
1: a 6.264538 15.1178117 3 4
2: a 1.836433 3.8984324 5 1
3: a 8.356286 6.2124058 2 2
4: a 15.952808 22.1469989 1 5
5: a 3.295078 11.2493092 4 3
6: b 8.204684 0.4493361 1 2
7: b 4.874291 0.1619026 4 1
8: b 7.383247 9.4383621 2 5
9: b 5.757814 8.2122120 3 4
10: b 3.053884 5.9390132 5 3
If there are many columns which are to be ranked individually, some descending, some ascending, we can do this in two steps
# first rank all columns in descending order
cols_desc <- c("y")
dt[, paste0("rank_", cols_desc) := lapply(.SD, frankv, ties.method = "min", order = -1L),
.SDcols = cols_desc, by = x][]
# then rank all columns in ascending order
cols_asc <- c("z")
dt[, paste0("rank_", cols_asc) := lapply(.SD, frankv, ties.method = "min", order = +1L),
.SDcols = cols_asc, by = x][]
x y z rank_y rank_z
1: a 6.264538 15.1178117 3 4
2: a 1.836433 3.8984324 5 1
3: a 8.356286 6.2124058 2 2
4: a 15.952808 22.1469989 1 5
5: a 3.295078 11.2493092 4 3
6: b 8.204684 0.4493361 1 2
7: b 4.874291 0.1619026 4 1
8: b 7.383247 9.4383621 2 5
9: b 5.757814 8.2122120 3 4
10: b 3.053884 5.9390132 5 3

Get number of same individuals for different groups

I have a data set with individuals (ID) that can be part of more than one group.
Example:
library(data.table)
DT <- data.table(
ID = rep(1:5, c(3:1, 2:3)),
Group = c("A", "B", "C", "B",
"C", "A", "A", "C",
"A", "B", "C")
)
DT
# ID Group
# 1: 1 A
# 2: 1 B
# 3: 1 C
# 4: 2 B
# 5: 2 C
# 6: 3 A
# 7: 4 A
# 8: 4 C
# 9: 5 A
# 10: 5 B
# 11: 5 C
I want to know the sum of identical individuals for 2 groups.
The result should look like this:
Group.1 Group.2 Sum
A B 2
A C 3
B C 3
Where Sum indicates the number of individuals the two groups have in common.
Here's my version:
# size-1 IDs can't contribute; skip
DT[ , if (.N > 1)
# simplify = FALSE returns a list;
# transpose turns the 3-length list of 2-length vectors
# into a length-2 list of 3-length vectors (efficiently)
transpose(combn(Group, 2L, simplify = FALSE)), by = ID
][ , .(Sum = .N), keyby = .(Group.1 = V1, Group.2 = V2)]
With output:
# Group.1 Group.2 Sum
# 1: A B 2
# 2: A C 3
# 3: B C 3
As of version 1.9.8 (on CRAN 25 Nov 2016), data.table has gained the ability to do non-equi joins. So, a self non-equi join can be used:
library(data.table) # v1.9.8+
setDT(DT)[, Group:= factor(Group)]
DT[DT, on = .(ID, Group < Group), nomatch = 0L, .(ID, x.Group, i.Group)][
, .N, by = .(x.Group, i.Group)]
x.Group i.Group N
1: A B 2
2: A C 3
3: B C 3
Explanantion
The non-equi join on ID, Group < Group is a data.table version of combn() (but applied group-wise):
DT[DT, on = .(ID, Group < Group), nomatch = 0L, .(ID, x.Group, i.Group)]
ID x.Group i.Group
1: 1 A B
2: 1 A C
3: 1 B C
4: 2 B C
5: 4 A C
6: 5 A B
7: 5 A C
8: 5 B C
We self-join with the same dataset on 'ID', subset the rows where the 'Group' columns are different, get the nrows (.N), grouped by the 'Group' columns, sort the 'Group.1' and 'Group.2' columns by row using pmin/pmax and get the unique value of 'N'.
library(data.table)#v1.9.6+
DT[DT, on='ID', allow.cartesian=TRUE][Group!=i.Group, .N ,.(Group, i.Group)][,
list(Sum=unique(N)) ,.(Group.1=pmin(Group, i.Group), Group.2=pmax(Group, i.Group))]
# Group.1 Group.2 Sum
#1: A B 2
#2: A C 3
#3: B C 3
Or as mentioned in the comments by #MichaelChirico and #Frank, we can convert 'Group' to factor class, subset the rows based on as.integer(Group) < as.integer(i.Group), group by 'Group', 'i.Group' and get the nrow (.N)
DT[, Group:= factor(Group)]
DT[DT, on='ID', allow.cartesian=TRUE][as.integer(Group) < as.integer(i.Group), .N,
by = .(Group.1= Group, Group.2= i.Group)]
Great answers above.
Just an alternative using dplyr in case you, or someone else, is interested.
library(dplyr)
cmb = combn(unique(dt$Group),2)
data.frame(g1 = cmb[1,],
g2 = cmb[2,]) %>%
group_by(g1,g2) %>%
summarise(l=length(intersect(DT[DT$Group==g1,]$ID,
DT[DT$Group==g2,]$ID)))
# g1 g2 l
# (fctr) (fctr) (int)
# 1 A B 2
# 2 A C 3
# 3 B C 3
yet another solution (base R):
tmp <- split(DT, DT[, 'Group'])
ans <- apply(combn(LETTERS[1 : 3], 2), 2, FUN = function(ind){
out <- length(intersect(tmp[[ind[1]]][, 1], tmp[[ind[2]]][, 1]))
c(group1 = ind[1], group2 = ind[2], sum_ = out)
}
)
data.frame(t(ans))
# group1 group2 sum_
#1 A B 2
#2 A C 3
#3 B C 3
first split data into list of groups, then for each unique pairwise combinations of two groups see how many subjects in common they have, using length(intersect(....

Summarize a dataframe by groups

Consider the following dataframe with 4 columns:
df = data.frame(A = rnorm(10), B = rnorm(10), C = rnorm(10), D = rnorm(10))
The columns A, B, C, D belong to different groups, and the groups are defined in a separate dataframe:
groups = data.frame(Class = c("A","B","C","D"), Group = c("G1", "G2", "G2", "G1"))
#> groups
# Class Group
#1 A G1
#2 B G2
#3 C G2
#4 D G1
I would like to average elements of the columns that belong to the same group, and get something similar to:
#> res
# G1 G2
#1 -0.30023039 -0.71075139
#2 0.53053443 -0.12397126
#3 0.21968567 -0.46916160
#4 -1.13775100 -0.61266026
#5 1.30388130 -0.28021734
#6 0.29275876 -0.03994522
#7 -0.09649998 0.59396983
#8 0.71334020 -0.29818438
#9 -0.29830924 -0.47094084
#10 -0.36102888 -0.40181739
where each cell of G1 is the mean of the relative cells of A and D, and each cell of G2 is the mean of the relative cells of B and C, etc.
I was able to achieve this result, but in a rather brute force way:
l = levels(groups$Group)
res = data.frame(matrix(nc = length(levels), nr = nrow(df)))
for(i in 1:length(l)) {
df.sub = df[which(groups$Group == l[i])]
res[,i] = apply(df.sub, 1, mean)
}
names(res) <- l
Is there a better way of doing this? In reality, I have more than 20 columns and more than 10 groups.
Thank you!
using data.table
library(data.table)
groups <- data.table(groups, key="Group")
DT <- data.table(df)
groups[, rowMeans(DT[, Class, with=FALSE]), by=Group][, setnames(as.data.table(matrix(V1, ncol=length(unique(Group)))), unique(Group))]
G1 G2
1: -0.13052091 -0.3667552
2: 1.17178729 -0.5496347
3: 0.23115841 0.8317714
4: 0.45209516 -1.2180895
5: -0.01861638 -0.4174929
6: -0.43156831 0.9008427
7: -0.64026238 0.1854066
8: 0.56225108 -0.3563087
9: -2.00405840 -0.4680040
10: 0.57608055 -0.6177605
# Also, make sure you have characters, not factors,
groups[, Class := as.character(Class)]
groups[, Group := as.character(Group)]
simple base:
tapply(groups$Class, groups$Group, function(X) rowMeans(df[, X]))
using sapply :
sapply(unique(groups$Group), function(X)
rowMeans(df[, groups[groups$Group==X, "Class"]]) )
I would personally go with Ricardo's solution, but another option would be to merge your two datasets first, and then use your preferred method of aggregating.
library(reshape2)
## Retain the "rownames" so we can aggregate by row
temp <- merge(cbind(id = rownames(df), melt(df)), groups,
by.x = "variable", by.y = "Class")
head(temp)
# variable id value Group
# 1 A 1 -0.6264538 G1
# 2 A 2 0.1836433 G1
# 3 A 3 -0.8356286 G1
# 4 A 4 1.5952808 G1
# 5 A 5 0.3295078 G1
# 6 A 6 -0.8204684 G1
## This is the perfect form for `dcast` to do its work
dcast(temp, id ~ Group, value.var="value", mean)
# id G1 G2
# 1 1 0.36611287 1.21537927
# 2 10 0.22889368 0.50592144
# 3 2 0.04042780 0.58598977
# 4 3 -0.22397850 -0.27333780
# 5 4 0.77073788 -2.10202579
# 6 5 -0.52377589 0.87237833
# 7 6 -0.61773147 -0.05053117
# 8 7 0.04656955 -0.08599288
# 9 8 0.33950565 -0.26345809
# 10 9 0.83790336 0.17153557
(Above data using set.seed(1) on your sample "df".

Resources