A local-search 2-approximation for 2-correlation-clustering

Abstract

CorrelationClustering is now an established problem in the algorithms and constrained clustering communities. With the requirement that at most two clusters be formed, the minimisation problem is related to the study of signed graphs in the social psychology community, and has applications in statistical mechanics and biological networks. Although a PTAS exists for this problem, its running time is impractical. We therefore introduce a number of new algorithms for 2CC, including two that incorporate some notion of local search. In particular, we show that the algorithm we call PASTA-toss is a 2-approximation on complete graphs. Experiments confirm the strong performance of the local search approaches, even on non-complete graphs, with running time significantly lower than rival approaches. © 2008 Springer-Verlag Berlin Heidelberg.