Efficient mining of distance-based subspace clusters

Traditional similarity measurements often become meaningless when dimensions of datasets increase. Subspace clustering has been proposed to find clusters embedded in subspaces of high dimensional datasets. Many existing algorithms use a grid based approach to partition the data space into nonoverlapping rectangle cells, and then identify connected dense cells as clusters. The rigid boundaries of the grid based approach may cause a real cluster to be divided into several small clusters. In this paper, we propose to use a sliding window approach to partition the dimensions to preserve significant clusters. We call this model nCluster model. The sliding window approach generates more bins than the gridbased approach, thus it incurs higher mining cost. We develop a deterministic algorithm, called MaxnCluster, to mine nClusters efficiently. MaxnCluster uses several techniques to speed up the mining, and it produces only maximal nClusters to reduce result size. Non-maximal nClusters are pruned without the need of storing the discovered nClusters in the memory, which is key to the efficiency of MaxnCluster. Our experiment results show that MaxnCluster can produce maximal nClusters efficiently and accurately.