Difference in timing for vectorized simple operations between an R matrix and a data.frame.

dataframe_matrix_timing.R

# Timing measurement in R. Vectorized operation on matrix / data.frame
# Author: José María Mateos - jmmateos@ieee.org
#
# For certain vectorized operations, it makes sense to convert your data.frame
# into a matrix. Even if you are using apply, the data frame iteration can be
# real slow.
#
# In this example, I will compute the Euclidean distance for a random vector
# and a random matrix / data.frame of thousands of elements. Operations will be
# done in two different ways: using the apply function over the columns and 
# directly using vectorized operations. The intuition says that the latter 
# should be faster, but as this example shows, this is not always the case.

#### GENERATE DATA ####
mm <- replicate(15000, rnorm(50))
md <- as.data.frame(mm)
# The vector whose distance to each other I want to measure
v <- rnorm(50)

#### MATRIX OPERATIONS ####
cat("Matrix time:\n")
# Method 1: apply
distance <- function(a, b) sqrt(sum((a - b)^2))
cat(" * With apply\n")
print(system.time(d1 <- apply(mm, 2, function(x) distance(x, v))))
# Method 2: direct vectorized operations
cat(" * Vectorized operation\n")
print(system.time(d2 <- sqrt(colSums((mm - v)^2))))
cat(paste("Are the two results identical?:", identical(d1, d2), "\n"))

#### DATA FRAME OPERATIONS ####
cat("data.frame time:\n")
# Method 1: apply
cat(" * With apply\n")
print(system.time(d1 <- apply(md, 2, function(x) distance(x, v))))
# Method 2: direct vectorized operations
cat(" * Vectorized operation\n")
print(system.time(d2 <- sqrt(colSums((md - v)^2))))
cat(paste("Are the two results identical?:", identical(d1, d2), "\n"))

My own results, in case anyone wants to check this but is too lazy to run the code:

Matrix time:
 * With apply
   user  system elapsed 
    0.2     0.0     0.2 
 * Vectorized operation
   user  system elapsed 
  0.004   0.008   0.014 
Are the two results identical?: TRUE 
data.frame time:
 * With apply
   user  system elapsed 
  0.504   0.000   0.503 
 * Vectorized operation
   user  system elapsed 
  6.008   0.036   6.055 
Are the two results identical?: TRUE