Subspace clustering through attribute clustering

Abstract

Many recently proposed subspace clustering methods suffer from two severe problems. First, the algorithms typically scale exponentially with the data dimensionality or the subspace dimensionality of clusters. Second, the clustering results are often sensitive to input parameters. In this paper, a fast algorithm of subspace clustering using attribute clustering is proposed to overcome these limitations. This algorithm first filters out redundant attributes by computing the Gini coefficient. To evaluate the correlation of every two non-redundant attributes, the relation matrix of non-redundant attributes is constructed based on the relation function of two dimensional united Gini coefficients. After applying an overlapping clustering algorithm on the relation matrix, the candidate of all interesting subspaces is achieved. Finally, all subspace clusters can be derived by clustering on interesting subspaces. Experiments on both synthesis and real datasets show that the new algorithm not only achieves a significant gain of runtime and quality to find subspace clusters, but also is insensitive to input parameters.

Keywords

subspace clustering, high dimensional data, attribute clustering