Data clustering is an important theoretical topic and a sharp tool for various applications. It is a task frequently arising in geometric computing. The main objective of data clustering is to partition a given data set into clusters such that the data items within the same cluster are "more" similar to each other with respect to certain measures. In this paper, we study the pairwise data clustering problem with pairwise similarity/dissimilarity measures that need not satisfy the triangle inequality. By using a criterion, called the minimum normalized cut, we model the general pairwise data clustering problem as a graph partition problem. The graph partition problem based on minimizing the normalized cut is known to be NP-hard. For an undirected weighted graph of n vertices, we present a ((4+o(1)) In n)-approximation polynomial time algorithm for the minimum normalized cut problem; this is the first provably good approximation polynomial time algorithm for the problem. We also give a more efficient algorithm for this problem by sacrificing the approximation ratio slightly. Further, our scheme achieves a ((2+o(1)) In n)-approximation polynomial time algorithm for computing the sparsest cuts in edge-weighted and vertex-weighted undirected graphs, improving the previously best known approximation ratio by a constant factor. Some applications and implementation work of our approximation normalized cut algorithms are also discussed.
A Polynomial-Time Algorithm for Minimizing the Deep Coalescence Cost for Level-1 Species Networks
