Evolving computationally efficient hashing for similarity search

Abstract

Finding nearest neighbors in high-dimensional spaces is a very expensive task. Locality-sensitive hashing is a general dimension reduction technique that maps similar elements closely in the hash space, streamlining near neighbor lookup. In this paper we propose a variable genome length biased random key genetic algorithm whose encoding facilitates the exploration of locality-sensitive hash functions that only use sparsely applied addition operations instead of the usual costly dense multiplications. Experimental results show that the proposed method obtains highly efficient functions with a much higher mean average precision than standard methods using random projections, while also being much faster to compute.