Informatica

Similarity searching has become widely available in many on-line archives of multimedia data. Users accessing such systems look for data items similar to their specific query object and typically refine results by re-running the search with a query from the results. We study this issue and propose a mechanism of approximate kNN query evaluation that incorporates statistics of accessing index data partitions. Apart from the distance between database objects, it also considers the prior query answers to prioritize index partitions containing frequently retrieved data, so evaluating repetitive similar queries more efficiently. We verify this concept in a number of experiments.