An exact CP approach for the cardinality-constrained euclidean minimum sum-of-squares clustering problem

Clustering consists in finding hidden groups from unlabeled data which are as homogeneous and well-separated as possible. Some contexts impose constraints on the clustering solutions such as restrictions on the size of each cluster, known as cardinality-constrained clustering. In this work we present an exact approach to solve the Cardinality-Constrained Euclidean Minimum Sum-of-Squares Clustering Problem. We take advantage of the structure of the problem to improve several aspects of previous constraint programming approaches: lower bounds, domain filtering, and branching. Computational experiments on benchmark instances taken from the literature confirm that our approach improves our solving capability over previously-proposed exact methods for this problem.