Performance of KDB-trees with query-based splitting

While the persistent data of many advanced database applications, such as OLAP and scientific studies, are characterized by very high dimensionality, typical queries posed on these data appeal to a small number of relevant dimensions. Unfortunately, the multidimensional access methods designed for high-dimensional data perform rather poorly for these partially specified queries. A potentially very appealing idea, frequently suggested in the literature, is to adopt a node-splitting policy that takes into account the "importance" of individual dimensions, which could be determined either a priori or through a statistical sampling of actual queries. This paper presents the results of some carefully controlled experiments conducted to observe the effects of query-based splitting on the performance of KDB-trees. The strategy is compared to a splitting policy that selects the split dimensions in a "cyclic" fashion, which has been shown to be very effective, especially in high-dimensional situations. Based on the results, the query-based splitting does not appear to be a very appealing splitting strategy for KDB-trees.