Cluster ensembles have attracted a lot of research interests in recent years, and their applications continue to expand. Among the various algorithms for cluster ensembles, those based on coassociation matrices are probably the ones studied and used the most because coassociation matrices are easy to understand and implement. However, the main limitation of coassociation matrices as the data structure for combining multiple clusterings is the complexity that is at least quadratic to the number of patterns N. In this paper, we propose CA-tree, which is a dendogram-like hierarchical data structure, to facilitate efficient and scalable cluster ensembles for coassociation-matrix-based algorithms. All the properties of the CA-tree are derived from base cluster labels and do not require the access to the original data features. We then apply a threshold to the CA-tree to obtain a set of nodes, which are then used in place of the original patterns for ensemble-clustering algorithms. The experiments demonstrate that the complexity for coassociation-based cluster ensembles can be reduced to close to linear to N with minimal loss on clustering accuracy.
|Number of pages||13|
|Journal||IEEE Transactions on Systems, Man, and Cybernetics, Part B: Cybernetics|
|State||Published - 1 Jun 2011|
- Cluster ensemble
- coassociation matrix
- multiple clusterings