Orthogonal Clustering

摘要

Often there are several logical criteria for clustering data. If all such clusterings are to be elicited for the purposes of cluster analysis it is typically necessary to use more than one algorithm. We have developed a novel clustering algorithm that uses a clustering generated by a previous pass of the algorithm as an additional feature to be optimized. The result is a new clustering that is distinct from the original clustering. Our method is analogous to boosting, except that where boosting converges, using a collection of models to form a single model, our method diverges, using a single model to generate one or more new models. We present experimental results using both synthetic data of our own design and real-world data found in the literature.