Subspace Clustering Ensembles through Tensor Decomposition

摘要

In recent years many different subspace clusteringalgorithms and related methods have been proposed. Theypromise to not only find hidden structures in data sets, but also toselect for each structure the features, which are most prominent. Yet, most of these methods suffer from the same problem:finding a satisfactory clustering result heavily depends on anadequate configuration of the parameters. In case of insufficientparameterization, a result is potentially hard to interpret andmight contain hundreds of clusters. For traditional clustering al-gorithms different ensemble methods have been developed, whichmitigate these effects by incorporating multiple clustering outputsinto a consensus result. However, most of these methods cannotbe straightforwardly adopted to include subspace information. We propose a novel subspace clustering ensemble algorithmSubCluEns based on the minimum description length principle. It allows combining multiple results of subspace and projectedclustering algorithms into a consensus clustering.