CODEWITHSUNDEEP
rdataframetime-series
Mike Furlender
Mike Furlender

Reputation: 4029

R: How to speed up this function?

I have a large data frame (named z) that looks like this:

    RPos    M1
    1   -0.00020
    2   0.00010
    3   -0.00012
    4   -0.00035
    5   -0.00038 
...etc (about 300,000 observations)

It is essentially a time series (although it is actually a data frame, not ts or zoo). Where RPos is the index number (explicitly stored), and M1 is any metric.

I have another data frame (named actionlist) with about 30,000 *non-consecutive observations. Each value in actionlist's RPos column represents the last of 34 consecutive points.

My final piece of data is a single data frame (named x) of only 34 consecutive observations.

My goal is to calculate the correlation coefficients between x and each observation in actionlist (which, again, is the end-point of 34 consecutive observations).

To do this I must generate these 34-point consecutive point time series segments from z (the large data frame).

Currently, I am doing it like this:

n1<-33:0
for(i in 1:nrow(actionlist))
{
    crs[i,2]<-cor(z[actionlist$RPos[i]+n1,2],x[,2])  
}

When looking at the Rprof readout this is what I get:

$by.self
              self.time self.pct total.time total.pct
[.data.frame       0.68    25.37       0.98     36.57
.Call              0.22     8.21       0.22      8.21
cor                0.16     5.97       2.30     85.82
...etc

It looks as though [.data.frame is taking the longest. Specifically I am pretty sure that it is this part: z[actionlist$RPos[i]+n1,2]

How can I speed up (eliminate the need for?) this part of the function?

I asked a similar question before, except instead of looking within a restricted list (actionlist) I was looking through every possible consecutive 34-observation within z. The answer was posted here, but I cannot figure out how to adapt it to a restricted list.

Any help would be very appreciated!

Upvotes: 1

Views: 319

Answers (1)

Vincent Zoonekynd
Vincent Zoonekynd

Reputation: 32401

The most straightforward is probably to build a matrix containing the data you want to compute the correlation with, and eschew the loop altogether.

# Sample data
n <- 3e5
m <- 3e4
k <- 35
z <- data.frame(
  RPos = 1:n,
  M1   = rnorm(n)
)
actionlist <- sample( k:n, m )
x <- rnorm(k)

system.time( for (j in 1:10) {
  # Index of the observations we want
  i <- sapply( (k-1):0, function(u) actionlist - u )
  # Data we want to compute the correlation with
  y <- matrix( z$M1[i], nr=nrow(i) )
  # Computations
  result <- cor(t(y),x)
} ) # 150ms per iteration

Upvotes: 4

Related Questions