kevinmcguinness
commented
6 years ago @Paul0629 The ranking is done via sparse matrix-vector multiplication, which is equivalent to an inverted index lookup in terms of computational complexity. It is straightforward to parallelize this computation when dealing with larger datasets: just split the dataset over multiple nodes, have each node compute an independent ranking (sparse matvec), and merge the resulting ranked lists.
@evamohe Dear Eva, BLCF is efficient in terms of storage and computation. But if the dataset is huge like 100K 0r 1M images, do we still traverse the whole dataset to rank them all. How about the inverted index or hash code or something. @xavigiro