Density and Non-Grid based Subspace Clustering via Kernel Density Estimation

Instead of finding clusters in full data space, subspace clustering has been an emergent task which aims at discovering clusters embedded in subspaces. Most of the previous works in the literature are grid and density based approaches, where the performance of clustering algorithms heavily depends on the partition of grids which may incur the serious loss of the real data distribution. In order to address the dilemma of grid partition, in this paper we propose a Density and Non-Grid based Subspace Clustering (DNGSC) algorithm via Kernel Density Estimation, which is able to discover arbitrarily shaped subspace clusters and insensitive to the only input parameter. The Kernel Density Estimation method with its firm mathematical basis has the ability to accurately uncover the underlying boundaries of data without any presumed canonical distribution. Besides, in order to deal with “the density divergence problem” which means the region densities vary in difference subspace cardinalities, different density thresholds for different subspaces are automatically determined to effectively discover arbitrarily shaped clusters in all subspaces, and FP-tree structure is employed to mine all dense subspace clusters with two efficient pruning strategies. To demonstrate the effectiveness and efficiency of DNGSC, we conduct extensive experiments on both synthetic and real data sets, and the performance comparison on accuracy and efficiency with DENCOS, MAFIA and CLIQUE shows the superiority of our new algorithm.