Efficient Cluster Detection by Ordered Neighborhoods

Detecting cluster structures seems to be a simple task, i.e. separating similar from dissimilar objects. However, given today's complex data, (dis-)similarity measures and traditional clustering algorithms are not reliable in separating clusters from each other. For example, when too many dimensions are considered simultaneously, objects become unique and (dis-) similarity does not provide meaningful information to detect clusters anymore. While the (dis-) similarity measures might be meaningful for individual dimensions, algorithms fail to combine this information for cluster detection. In particular, it is the severe issue of a combinatorial search space that results in inefficient algorithms. In this paper we propose a cluster detection method based on the ordered neighborhoods. By considering such ordered neighborhoods in each dimension individually, we derive properties that allow us to detect clustered objects in dimensions in linear time. Our algorithm exploits the ordered neighborhoods in order to find both the similar objects and the dimensions in which these objects show high similarity. Evaluation results show that our method is scalable with both database size and dimensionality and enhances cluster detection w.r.t. state-of-the-art clustering techniques.