A Dynamic Pivot Selection Technique for Similarity Search

All pivot-based algorithms for similarity search use a set of reference points called pivots. The pivot-based search algorithm precomputes some distances to these reference points, which are used to discard objects during a search without comparing them directly with the query. Though most of the algorithms proposed to date select these reference points at random, previous works have shown the importance of intelligently selecting these points for the index performance. However, the proposed pivot selection techniques need to know beforehand the complete database to obtain good results, which inevitably makes the index static. More recent works have addressed this problem, proposing techniques that dynamically select pivots as the database grows. This paper presents a new technique for choosing pivots, that combines the good properties of previous proposals with the recently proposed dynamic selection. The experimental evaluation provided in this paper shows that the new proposed technique outperforms the state-of-art methods for selecting pivots.