Improved Densification of One Permutation Hashing
This improves near neighbor search efficiency for sparse web data, but it is incremental as it builds on existing densification methods.
The paper tackled the sub-optimal densification in one permutation hashing for LSH, which affected accuracy on sparse datasets, and proposed a new procedure that is provably better, with experimental support on public datasets.
The existing work on densification of one permutation hashing reduces the query processing cost of the $(K,L)$-parameterized Locality Sensitive Hashing (LSH) algorithm with minwise hashing, from $O(dKL)$ to merely $O(d + KL)$, where $d$ is the number of nonzeros of the data vector, $K$ is the number of hashes in each hash table, and $L$ is the number of hash tables. While that is a substantial improvement, our analysis reveals that the existing densification scheme is sub-optimal. In particular, there is no enough randomness in that procedure, which affects its accuracy on very sparse datasets. In this paper, we provide a new densification procedure which is provably better than the existing scheme. This improvement is more significant for very sparse datasets which are common over the web. The improved technique has the same cost of $O(d + KL)$ for query processing, thereby making it strictly preferable over the existing procedure. Experimental evaluations on public datasets, in the task of hashing based near neighbor search, support our theoretical findings.