ASCRClu: an adaptive subspace combination and reduction algorithm for clustering of high-dimensional data

The curse of dimensionality in high-dimensional data is one of the major challenges in data clustering. Recently, a considerable amount of literature has been published on subspace clustering to address this challenge. The main objective of the subspace clustering is to discover clusters embedded in any possible combination of the attributes. Previous studies have mostly been generating redundant subspace clusters, leading to clustering accuracy loss and also increasing the running time. In this paper, a bottom-up density-based approach is proposed for clustering of high-dimensional data. We employ the cluster structure as a similarity measure to generate the optimal subspaces which result in raising the accuracy of the subspace clustering. Using this idea, we propose an iterative algorithm to discover similar subspaces using the similarity in the features of subspaces. At each iteration of this algorithm, it first determines similar subspaces, then combines them to generate higher-dimensional subspaces, and finally re-clusters the subspaces. The algorithm repeats these steps and converges to the final clusters. Experiments on various synthetic and real datasets show that the results of the proposed approach are significantly better in both quality and runtime comparing to the state of the art on clustering high-dimensional data. The accuracy of the proposed method is around 34% higher than the CLIQUE algorithm and around 6% higher than DiSH.