Shrinking data balls in metric indexes

Bilegsaikhan Naidan, Magnus Lie Hetland
unpublished
Some of the existing techniques for approximate similarity retrieval in metric spaces are focused on shrinking the query region by user-defined parameter. We modify this approach slightly and present a new approximation technique that shrinks data regions instead. The proposed technique can be applied to any metric indexing structure based on the ball-partitioning principle. Experiments show that our technique performs better than the relative error approximation and region proximity
more » ... and that it achieves significant speedup over exact search with a low degree of error. Beyond introducing this new method, we also point out and remedy a problem in the relative error approximation technique, substantially improving its performance.
fatcat:e3pleaunjndujadynciucyducm