BM-index: Balanced Metric Space Index Based on Weighted Voronoi Partitioning

Abstract

Processing large volumes of various data requires index structures that can efficiently organize them on secondary memory. Methods based on pivot permutations have become popular because of their tremendous querying performance. Pivot permutations can be perceived as a recursive Voronoi tessellation with a fixed set of anchors. Its disadvantage is that it cannot adapt to the data distribution well, which leads to cells unbalanced in occupation and unevenly filled disk buckets. In this paper, we address this issue and propose a novel schema called the BM-index. It exploits a weighted Voronoi partitioning, which is able to respect the data distribution. We present an algorithm to balance the data partitions, and show its correctness. The secondary memory is then accessed efficiently, which is shown in experiments executing k-nearest neighbors queries on a real-life image collection CoPhIR.