Abstract—This paper proposes a novel constrained clustering method that is based on a graph-cut problem formalized by SDP (Semi-Definite Programming). Our SDP approach has the advantage of convenient constraint utilization compared with conventional spectral clustering methods. The algorithm starts from a single cluster of a whole dataset and repeatedly selects the largest cluster, which it then divides into two clusters by swapping rows and columns of a relational label matrix obtained by solving the maximum graph-cut problem. This swapping procedure is effective because we can create clusters without any computationally heavy matrix decomposition process to obtain a cluster label for each data. The results of experiments using datasets from the ODP and WebKB corpus demonstrated that our method outperformed other conventional and the state of the art clustering methods in many cases. In particular, we discuss the difference between our approach and another similar one that uses the same SDP formalization as ours. Since the number of constraints used in the experiments is relatively small and can be practical for human feedback, we consider our clustering provides a promising basic method to interactive Web clustering.  

Cite: Masayuki Okabe and Seiji Yamada, "Graph-cut based Constrained Clustering by Grouping Relational Labels," Journal of Emerging Technologies in Web Intelligence, Vol. 4, No. 1, pp. 43-50, February 2012. doi:10.4304/jetwi.4.1.43-50