KL-Divergence Implementation does not handle 0 probabilities

[carlosparadis]

When executing createJSON, the following error will be thrown:

Error in stats::cmdscale(dist.mat, k = 2) : NA values not allowed in 'd'

I traced it down to:

LDAvis/R/createJSON.R

Lines 298 to 304 in 51bb51e

	# first, we compute a pairwise distance between topic distributions
	# using a symmetric version of KL-divergence
	# http://en.wikipedia.org/wiki/Jensen%E2%80%93Shannon_divergence
	jensenShannon <- function(x, y) {
	m <- 0.5*(x + y)
	0.5*sum(x*log(x/m)) + 0.5*sum(y*log(y/m))
	}

To reproduce the issue:

Reproducible dataset

x <- c(0.2,0.3,0.3)
y <- c(0.2,0.3,0.4) 
b <- c(0.2,0.3,0) 

Using LDAvis implementation shown at the start of this issue:

> jensenShannon(x=x,y=y)
[1] 0.003583677
> jensenShannon(x=x,y=b)
[1] NaN

The same test, using cosine function from lsa package:

> cosine(x=x,y=y)
          [,1]
[1,] 0.9897595
> cosine(x=x,y=b)
          [,1]
[1,] 0.7687061

[rnkazman]

This seems like an implementation detail, not a principled reason to use one or the other. Is that correct?

[cpsievert]

Done in c7234d7

[caitsimop]

Hi there, I'm still getting this error in v0.3.5. Is this the most up to date version?