KSR-Tree: A Clustering Based High-Dimensional Indexing Approach

Abstract

In high-dimensional index, overflow split has been verified to be critical to the performance of kNN query processing. A node is split into two parts in the traditional way, however, these methods tend to result in many overlap regions in high-dimensional spaces which will significantly degrade the performance of retrieval. In this paper, we propose a method named KSR-Tree, it making use of a clustering based split algorithm to divides the node into multiple parts, and most of overlap regions will guarantee to be placed into the same node. This approach not only increased the capacity for newly arrived records, but also decreases the splitting overhead and reduces the overlap regions, thus the frequency of node splitting will reduced and meanwhile the retrieval performance obtains improvement. In the experiments, our results showed that the performance KSR-Tree significantly improved the performance of kNN query processing.