A coarse-grain grid-based subspace clustering method for online multi-dimensional data streams

This paper proposes a subspace clustering algorithm which combines grid-based clustering with frequent itemset mining. Given a d-dimensional data stream, the on-going distribution statistics of its data elements in every one-dimensional data space is monitored by a list of fine-grain grid-cells called a sibling list, so that all the one-dimensional clusters are accurately identified. By tracing a set of frequently co-occurred one-dimensional clusters, it is possible to find a coarse-grain dense rectangular space in a higher dimensional subspace. An ST-tree is introduced to continuously monitor dense rectangular spaces in all the subspaces of the d dimensions. Among the spaces, those ones whose densities are greater than or equal to a user defined minimum support threshold S_min are corresponding to final clusters.