K-maximin clustering: a maximin correlation approach to partition-based clustering

We propose a new clustering algorithm based upon the maximin correlation analysis (MCA), a learning technique that can minimize the maximum misclassification risk. The proposed algorithm resembles conventional partition clustering algorithms such as k-means in that data objects are partitioned into k disjoint partitions. On the other hand, the proposed approach is unique in that an MCA-based approach is used to decide the location of the representative point for each partition. We test the proposed technique with typography data and show our approach outperforms the popular k-means and k-medoids clustering in terms of retrieving the inherent cluster membership.