We consider a semi-supervised clustering problem, where selected pairs of data points are labeled by an expert as must-links or cannot-links. Basically, must-link constraints indicate that two points should be grouped together, while those with cannot-link constraints should be grouped separately. We present a clustering algorithm, which creates a partition consistent with pairwise constraints by maximizing the probability of correct assignments. Moreover, unlabeled data are used by maximizing their prediction confidence. Preliminary experimental studies show that the proposed method gives accurate results on sample data sets. Moreover, its kernelization allows to discover clustering patterns of arbitrary shapes.
Archived Files and Locations
|application/pdf 222.9 kB ||