Classification with Multi-Modal Classes Using Evolutionary Algorithms and Constrained Clustering

Abstract

Constrained clustering has been an active research topic in the last decade. Among the different kinds of constraints, must-link and cannot-link are the most adopted ones. However, most algorithms assume that the number of clusters are known a priori. Besides this usually unrealistic assumption, one often ignores the fact that must-link constraints may correspond to objects in different density regions of the input space, thereby requiring a more complex structure to represent the underlying concept. Aimed at overcoming these limitations, we present the Feasible-Infeasible Evolutionary Create & Eliminate for Expectation Maximization (FIECE-EM), which identifies a Gaussian Mixture Model that is a good fit for the data, while meeting the constraints provided. We compare FIECE-EM with a state-of-the-art algorithm. Our results indicate that FIECE-EM obtains competitive results, without the need for fine-tuning a tradeoff parameter as in the state-of-the-art algorithm under comparison.