Storing Data Once in M-tree and PM-tree

Abstract

Since the publication of M-tree, several enhancements were proposed to its structure. One of the most exciting is the use of additional global pivots that resulted in the PM-tree. In this paper, we revisit both M-tree and PM-tree to propose a new construction algorithm that stores data elements once in their trees hierarchies. The main challenge is to select data elements when an inner node split is needed. The idea is that as a data element is evaluated for pruning during traversal, it can become part of the result set, allowing faster convergence of nearest neighbor algorithms. The new insert and query algorithms enable faster retrieval, the decrease in node occupation of trees built with the same parameters, and also a reduction in the overlap among nodes, as shown in the experimental evaluation.