minor bugfixes & optimizations

[0xTiger]

Interesting paper :)

There were a few changes I needed to make to get this to work. The most notable being that converting to a sparse.coo_matrix and then acessing row & column indicies is much more memory efficient than converting to a sparse.dok_matrix. In fact for a reasonably large matrix the sparse.dok_matrix did not fit in memory for me. Aside from that they are equivalent.