Towards Artificial Priority Queues for Similarity Query Execution

Content-based retrieval in large collections of unstructured data is challenging not only from the difficulty of the defining similarity between data images where the phenomenon of semantic gap appears, but also the efficiency of execution of similarity queries. Search engines providing similarity search typically organize various multimedia data, e.g. images of a photo stock, and support k-nearest neighbor query. Users accessing such systems then look for data items similar to their specific query object and refine results by re-running the search with an object from the previous query results. This paper is motivated by unsatisfactory query execution performance of indexing structures that use metric space as a convenient data model. We present performance behavior of two state-of-the-art representatives and propose a new universal technique for ordering priority queue of data partitions to be accessed during kNN query evaluation. We verify it in experiments on real-life data-sets.