A Resourceful Algorithm for High Dimensional Spaces
Aretty Narayana . · 2013
The approaches for resemblance search in correlated, high dimensional data sets are derived with in a clustering framework. To conflict the “Curse of Dimensionality” a technique namely Vector approximation by indexing was proposed, employs scalar quantization and necessarily ignores dependencies across dimensions and a source of sub optimality is represented. In contrast inter dimensional correlations are exploited by clustering and thus a more packed in representation of data set. However, existing methods reduce the irrelevant clusters which are based on bounding hyper spheres and bounding rectangles which lack their tightness and compromises the efficiency in nearest neighbor search. In this paper, based on separating hyper plane boundaries of Voronoi clusters we propose a new cluster adaptive distance bound to complement our cluster based index. With a relatively small preprocessing storage overhead a bound is enabled by an efficient spatial filtering and is applicable to similarity measures. Experiments show that our indexing method is scalable with data set size and dimensionality and outperforms with several recent proposed indexes.