Efficient Similarity Search by Combining Indexing and Caching Strategies?

Abstract. A critical issue in large scale search engines is to efficiently handle sudden peaks of incoming query traffic. Research in metric spaces has addressed this problem from the point of view of creating caches that provide information to, if possible, exactly/approximately answer a query very quickly without needing to further process an index. However, one of the problems of that approach is that, if the cache is not able to provide an answer, the distances computed up to that moment are wasted, and the search must proceed through the index structure. In this paper we present an index structure that serves a twofold role: that of a cache and an index in the same structure. In this way, if we are not able to provide a quick approximate answer for the query, the distances computed up to that moment are used to query the index. We present an experimental evaluation of the performance obtained with our structure.