An adaptive kernel method for semi-supervised clustering

Abstract

Semi-supervised clustering uses the limited background knowledge to aid unsupervised clustering algorithms. Recently, a kernel method for semi-supervised clustering has been introduced, which has been shown to outperform previous semi-supervised clustering approaches. However, the setting of the kernel's parameter is left to manual tuning, and the chosen value can largely affect the quality of the results. Thus, the selection of kernel's parameters remains a critical and open problem when only limited supervision, provided in terms of pairwise constraints, is available. In this paper, we derive a new optimization criterion to automatically determine the optimal parameter of an RBF kernel, directly from the data and the given constraints. Our approach integrates the constraints into the clustering objective function, and optimizes the parameter of a Gaussian kernel iteratively during the clustering process. Our experimental comparisons and results with simulated and real data clearly demonstrate the effectiveness and advantages of the proposed algorithm. © Springer-Verlag Berlin Heidelberg 2006.