Sharing Hash Codes for Multiple Purposes

• DOI: 10.48550/arxiv.1609.03219
• Publication Date: 2016-01-01
• Authors: Wikor Pronobis; Danny Panknin; Johannes Kirschnick; Vignesh Srinivasan; Wojciech Samek; Volker Markl; Manohar Kaul; Klaus‐Robert Mueller; Shinichi Nakajima
• Keywords: Locality-sensitive hashing; Dynamic perfect hashing; Hamming distance; Sublinear function

Locality sensitive hashing (LSH) is a powerful tool for sublinear-time approximate nearest neighbor search, and variety of schemes have been proposed different dissimilarity measures. However, hash codes significantly depend on the dissimilarity, which prohibits users from adjusting at query time. In this paper, we propose {multiple purpose LSH (mp-LSH) shares dissimilarities. mp-LSH supports L2, cosine, inner product dissimilarities, their corresponding weighted sums, where weights can be adjusted It also allows us to modify importance pre-defined groups features. Thus, enables us, example, retrieve similar items with user preference taken into account, find material some properties (stability, utility, etc.) optimized, turn or off part multi-modal information (brightness, color, audio, text, in image/video retrieval. We theoretically empirically analyze performance three variants mp-LSH, demonstrate usefulness real-world data sets.