Performance Study of Independent Anchor Spaces for Similarity Searching

This work targets the problem of search efficiency vs. answer quality of approximate metric-based similarity search. We especially focus on techniques based on recursive Voronoi-like partitioning or, from another perspective, on pivot permutations. These techniques use sets of reference objects (anchors/pivots) to partition the metric space into cells of close data items. Instead of refining the search space by enlarging the anchor set of a single index, we propose to divide a large pivot set into several subsets and build multiple indexes with independent space partitioning; at query time, the overall search costs are also divided among the separate indexes. Our thorough experimental study on three different real datasets uncovers drawbacks of excessive increase of a single pivot set size—such partitioning refinement can be counterproductive beyond a certain number of pivots. Our approach overcomes the root causes of this limitation and increases the answer quality while preserving the search costs. Further, we address the question of robustness of the answer quality, which can be significantly improved by utilization of independent anchor spaces.