Bulk loading the MKL-tree

Abstract

MKL-tree is a hierarchical, height-balanced structure for high dimensional data indexing. This structure is based on data representation in a lower dimensional space by means of the MKL transform, a multi-space generalization of the KL transform. A local dimensionality reduction is performed at each node of the tree, allowing more selective features to be extracted and thus increasing the discriminating power of the index. The dynamical version of MKL-tree presents two main drawbacks: first, the incremental loading of data points can determine very different structures and, as a consequence, different query performance, depending on the insertion order; second, the creation of the index can be very expensive, due to the high number of updating required. Since, in real applications, a large dataset is usually available at the tree creation time, we propose a new bulk loading technique for MKL-tree, based on a recursive clustering of data objects. The new algorithm searches for an optimal partitioning of data points, in order to calculate the most suitable KL-subspaces to represent the dataset. Experimental results show that bulk loading can significantly improve the index performance with respect to the incremental insertion procedure, both in terms of effectiveness of similarity searches and of efficiency of the loading procedure. © Springer-Verlag Berlin Heidelberg 2003.