Why can cosine similarity between two vectors be negative?

Question

I have 2 vectors with 11 dimentions.

a <- c(-0.012813841, -0.024518383, -0.002765056,  0.079496744,  0.063928973,
        0.476156960,  0.122111977,  0.322930189,  0.400701256,  0.454048860,
        0.525526219)

b <- c(0.64175768,  0.54625694,  0.40728261,  0.24819750,  0.09406221, 
       0.16681692, -0.04211932, -0.07130129, -0.08182200, -0.08266852,
       -0.07215885)

cosine_sim <- cosine(a,b)

which returns:

-0.05397935

I used cosine() from lsa package.

for some values i am getting negative cosine_sim like the given one. I am not sure how the similarity can be negative. It should be between 0 and 1.

Can anyone explain what is going on here.

Answer 1

The nice thing about R is that you can often dig into the functions and see for yourself what is going on. If you type cosine (without any parentheses, arguments, etc.) then R prints out the body of the function. Poking through it (which takes some practice), you can see that there is a bunch of machinery for computing the pairwise similarities of the columns of the matrix (i.e., the bit wrapped in the if (is.matrix(x) && is.null(y)) condition, but the key line of the function is

crossprod(x, y)/sqrt(crossprod(x) * crossprod(y))

Let's pull this out and apply it to your example:

> crossprod(a,b)/sqrt(crossprod(a)*crossprod(b))
            [,1]
[1,] -0.05397935
> crossprod(a)
     [,1]
[1,]    1
> crossprod(b)
     [,1]
[1,]    1

So, you're using vectors that are already normalized, so you just have crossprod to look at. In your case this is equivalent to

> sum(a*b)
[1] -0.05397935

(for real matrix operations, crossprod is much more efficient than constructing the equivalent operation by hand).

As @Jack Maney's answer says, the dot product of two vectors (which is length(a)*length(b)*cos(a,b)) can be negative ...

For what it's worth, I suspect that the cosine function in lsa might be more easily/efficiently implemented for matrix arguments as as.dist(crossprod(x)) ...

edit: in comments on a now-deleted answer below, I suggested that the square of the cosine-distance measure might be appropriate if one wants a similarity measure on [0,1] -- this would be analogous to using the coefficient of determination (r^2) rather than the correlation coefficient (r) -- but that it might also be worth going back and thinking more carefully about the purpose/meaning of the similarity measures to be used ...