Fast similarity search in small-world networks

Abstract

We present a novel graph-based approach for fast similarity searches suitable for large-scale and high-dimensional data sets. We focus on a well-known feature of small-world networks, they are "searchable," and propose an efficient index structure called a degree-reduced nearest neighbor graph. A similarity search is then formulated as a problem of finding the most similar object to a query object by following the links in this graph with a best-first neighborhood search algorithm. The experimental results show that the proposed search method significantly reduces search costs. In particular, we apply it to data sets consisting of nearly one million documents, and successfully reduce the average number of similarity evaluations to only 0.9% of the total number of documents. © 2009 Springer-Verlag Berlin Heidelberg.