Hi everyone,

to put it succinctly, here's the BM25 equation:

f(w,D) * (k+1) / (k*B + f(w,D))

where w is the word, and D is the document (corresponding to rows and columns, respectively). f is a sparse matrix because only a fraction of the whole vocabulary of words appears in any given single document.

B is a function of only the document, but it doesn't matter, you can think of it as a constant if you want.

The problem is since f(w,D) is almost always zero, I only need to do the calculation (ie. multiply by (k+1) then divide by (k*B + f(w,D))) when f(w,D) is not zero. Is there a clever way to do this with masks?

You can refactor the above equation to get this:

(k+1)/(k*B/f(w,D) + 1) but alas we still have f(w,D) appearing in a denominator, which is bad (because of dividing by zero).

So anyway, currently I am converting to a coo_matrix and iterator through the non-zero values like this:

    cx = x.tocoo()    
    for i,j,v in itertools.izip(cx.row, cx.col, cx.data):
        (i,j,v)

That iterator is incredibly fast, but unfortunately coo_matrix does not support assignment. So I create a new copy of either a dok sparse matrix or a regular numpy array and assign to that.

I could also deal directly with the .data, .indptr, and indices attributes of csr_matrix, and see if it's possible to create a copy of .data attribute and update the values accordingly. I was hoping somebody had encountered this type of issue before.

Sincerely,

Basil Beirouti