Combining Clusterings With Different Detail Levels
Oded Kaminsky, Bar-Ilan University
Jacob Goldberger, Bar-Ilan University

Abstract:
In this study we address the problem of recovering a clustering of a dataset based on several clusterings provided by different experts. These experts provide clusterings on different levels (coarser or finer than the others). We present an automatic algorithm that combines the information provided by the experts into a single clustering that can be viewed as the average point of the input clusterings. We formulate the problem as an instance of correlation clustering and apply integer linear programming to obtain the average clustering. As a byproduct, we also obtain for each expert its reliability and the detail level encoded in its clustering. We apply the proposed algorithm to the task of averaging several image segmentations. The average segmentation is efficiently computed by first grouping the image into superpixels and then applying the proposed algorithm on the superpixel map. The performance of the proposed algorithm is demonstrated on manually annotated images from the Berkeley segmentation dataset.