Density-based subspace clustering over decayed windows on data streams

Some clusters usually hide in different dimensionality subspace. In order to tracing the subspace (dusters on line, a tree-like structure is proposed to contain the summarized information in this paper. And the density-based clustering method is applied. Initially, a partition for each dimension monitors its one-dimensional subclusters at the first level of the tree. A cell, the density of which exceeds our density threshold, is partitioned into several equal-size smaller cells. When a unit cell becomes dense, a set of new nodes are created as its child nodes. Here, we use different density thresholds for different height cells. A k-dimensional subspace cluster is found as a list of adjacent dense grid-cells at the kth height of the tree. Experimental results show that our algorithm is able to find all subspace clusters, even in environment with noises.