Compressing Locality Sensitive Hashing Tables

LSH is the industry standard for proximity searching tasks on collections of data having coordinates. An LSH index applies a set of hashing functions to the representation of an object to identify proximal objects to a query, leaving distal objects apart. In other words, objects with the same hash will be mutually proximal with high probability. LSH is very fast and gives probabilistic guarantees on the quality of the results. On the other hand, mobile applications using proximity queries are becoming common place. Feature extraction can be done in a smart phone. However, the actual query rely on a wireless link because memory is a scarce resource. To tackle the above problem, we present in this paper a method to compress the LSH index while still being able to query without decompressing. The query speed is practically the same, and can even be faster. We derive a lower bound on the memory requirements for the compress representation and present an implementation using close to optimal storage. We provide an extensive experimental comparison of our compressed representation against the uncompressed one over a large database of 55 million objects. We obtained a compression ratio ranging from 70% to 80% without slowing down, in practice, the search speed.