Cascaded cluster ensembles

Combining multiple clusterers is emerged as a powerful method for improving both the robustness and the stability of unsupervised classification solutions. In this paper, a framework for cascaded cluster ensembles is proposed, in which there are two layers of clustering. The first layer is considering about the diversity of clustering, and generating different partitions. In doing so, the samples in input space are mapped into labeled samples in a label attribute space whose dimensionality equals the ensemble size. In the second layer clustering, we choose a clustering algorithm as the consensus function. In other words, a combined partition is given by using the clustering algorithm on these labeled samples instead of input samples. In the second layer, we use the reduced k-means, or the reduced spectral, or the reduced hierarchical linkage algorithms as the clustering algorithm. For comparison, nine consensus functions, four of which belong to cascaded cluster ensembles are used in our experiments. Promising results are obtained for toy data as well as UCI data sets.