Me-Meoa/Dcc : Multiobjective Constrained Clustering Through Decomposition-Based Memetic Elitism

A B S T R A C T Traditionally unsupervised, clustering techniques have received renewed attention recently, as they have been shown to produce better results when provided with incomplete information about the dataset in the form of constraints. Combining classic clustering and constraints leads to constrained clustering, a semi-supervised learning problem still unexplored in many aspects. Based on the exploration-exploitation requirements of constrained clustering, a memetic elitist multiobjective evolutionary algorithm based on decomposition is proposed, which combines classic multiobjective optimization strategies with single-objective optimization procedures. The application scheme of our proposal for the constrained clustering problem is scrutinized and compared to several state-of-the-art methods for 20 datasets with incremental levels of constraint-based information. Experimental results, supported by Bayesian statistical testing, show a consistent improvement in clustering and multiobjective optimization related measures in favor of our proposal over the state-of-the-art.