Vectorizing function to calculate entropy

user_3373438 0
Viewed 286

I'm writing a function to calculate a Shannon diversity index. I have wide data with the pct of observations for each value as separate variables, with each row representing a different site. I will have between 2 and 7 variables depending on the data set. For each row I want to calculate the information index.

equation.

I have a loop function, but its quite slow and am looking for help to vectorize it. I'm also happy for a tidyverse style solution.

I've been looking at using the entropy package, but it seems to expect the data in long form, and while I could expand my data back out, that seems like it would be unnecessarily slow. I currently have 20k sites with 100's to 1000's of observations per site that have already been summarized into the wide format percent distributions. This question similarly works with long form data.

Example data

# Wide data, between 2 and 7 columns recording the percent of observations with each value, example using 3
df <- data.frame(
  site = 1:3, 
  l1 = c(.33, .5, 0), 
  l2 = c(.33, .5, 0), 
  l3 = c(.33, 0, 1)
)

Current loop function

entropy <- function(df, vars) {

  entropy_calc <- function(df, i, vars) {
    sum <- 0
    for (j in vars) { 
      x <- df[i,j]
      if(x != 0) { # skip zeros
        sum <- sum + x * log(x)
      }
    }
    return(-sum)
  }

  entropy <- rep(NA, nrow(df))
  for(i in 1:nrow(df)) {
    entropy[i] <- entropy_calc(df, i, vars)
  }

  return(as.numeric(entropy))
}

df$entropy <- entropy(df, 2:4)
vectorization r
1 Answers

This can be vectorized easily because the underlying functions needed are already vectorized. You don't need to manually skip zeroes because log(0) returns -Inf and 0*log(0) returns NaN. You can omit the NaN when summing the cell values by specifying na.rm = TRUE. 

entropy <- function(p) rowSums(-(p * log(p)), na.rm = TRUE)

entropy(df[,2:4])

Also check out the diversity() function in the vegan package which does essentially this, among other possibilities.

user_28546080
Related