A case for hubness removal in high-dimensional multimedia retrieval

Abstract

This work investigates the negative effects of hubness on multimedia retrieval systems. Because of a problem of measuring distances in high-dimensional spaces, hub objects are close to an exceptionally large part of the data while anti-hubs are far away from all other data points. In the case of similarity based retrieval, hub objects are retrieved over and over again while anti-hubs are nonexistent in the retrieval lists. We investigate textual, image and music data and show how re-scaling methods can avoid the problem and decisively improve the overall retrieval quality. The observations of this work suggest to make hubness analysis an integral part when building a retrieval system. © 2014 Springer International Publishing Switzerland.