Iterative K-Means Algorithm Based on Fisher Discriminant

K-Means clustering is a well-known tool in unsupervised learning. The performance of K-Means clustering, measured by the F-ratio validity index, highly depends on selection of its initial partition. This problematic dependency always leads to a local optimal solution for k-center clustering. To overcome this difficulty, we present an intuitive approach that iteratively incorporates Fisher discriminant analysis into the conventional K-Means clustering algorithm. In other words, at each time, a suboptimal initial partition for K-Means clustering is estimated by using dynamic programming in the discriminant subspace of input data. Experimental results show that the proposed algorithm outperforms the two comparative clustering algorithms, the PCA-based suboptimal K-Means clustering algorithm and the kd-tree based K-Means clustering algorithm.