Constrained clustering with k-means

We introduce a k−means type clustering in the presence of cannot–link and must–link constraints. First we apply a BIRCH type methodology to eliminate must–link constraints. Next we introduce a penalty function to substitute cannot–link constraints. When penalty values increase to +∞ the original cannot–link constraints are recovered. The preliminary numerical experiments show that constraints have the potential to significantly improve partitions generated by clustering algorithms. Keyword List clustering, k−means, must–links, cannot–links.