Indexing Dense Nested Metric Spaces for Efficient Similarity Search

Abstract. Searching in metric spaces is a very active field since it offers methods for indexing and searching by similarity in collections of unstructured data. These methods select some objects of the collection as reference objects to build the indexes. It has been shown that the way the references are selected affects the search performance, and several algorithms for good reference selection have been proposed. Most of them assume the space to have a reasonably regular distribution. However, in some spaces the objects are grouped in small dense clusters that can make these methods perform worse than a random selection. In this paper, we propose a new method able to detect these situations and adapt the structure of the index to them. Our experimental evaluation shows that our proposal is more efficient than previous approaches when using the same amount of memory.

• Type: Conference
• Year: 2009
• Where: ERSHOV
• Authors: Nieves R. Brisaboa, Miguel Rodríguez Luaces, Oscar Pedreira, Ángeles S. Places, Diego Seco
• Keywords: ERSHOV 2009, Metric Spaces, Reference Objects, Small Dense Clusters, Theoretical Computer Science