Excluded Middle Forest Versus Vantage Point Tree: An Analytical and Empirical Comparison

Abstract

A critical performance issue in searching distance-based indexing trees is that commonly multiple nodes, or data partitions, have to be further examined at each level of the tree. As a result, logarithmic search time is commonly not achievable. To solve this problem, excluded middle forest builds a tree for each data partition to be descended, so that logarithmic search time is achieved for each tree. Promising empirical results are reported from excluded middle forest. However, we observe that what determines whether a data partition has to be further examined is that data composing it, but not its location. Although each tree of the excluded middle forest can be searched in logarithmic time, the overall query performance is encumbered with the existence of multiple trees. We show analytically and empirically that multiple vantage point tree outperforms excluded middle forest.