Approximate similarity search: A multi-faceted problem

AbstractWe review the major paradigms for approximate similarity queries and propose a classification schema that easily allows existing approaches to be compared along several independent coordinates. Then, we discuss the impact that scheduling of index nodes can have on performance and show that, unlike exact similarity queries, no provable optimal scheduling strategy exists for approximate queries. On the positive side, we show that optimal-on-the-average schedules are well-defined and that their performance is indeed the best among practical schedules