Medoid-based data clustering with estimation of distribution algorithms
Henry E. L. Cagnini, Rodrigo Coelho Barros, Christian Vahl Quevedo, Márcio P. Basgalupp · 2016
Data clustering is the machine learning task that aims at arranging data into groups (clusters) of objects according to a similarity criterion. From an optimisation perspective, it is a particular kind of NP-hard grouping problem, thus attracting much attention from the evolutionary computation community. In this paper, we propose a novel data clustering algorithm based on a univariate estimation of distribution algorithm, namely Clus-EDA. It employs a medoid-based representation in which the cluster prototypes necessarily coincide with objects from the dataset. We compare Clus-EDA with both traditional non-evolutionary clustering algorithms such as k-means and hierarchical agglomerative clustering, and also with an evolutionary algorithm for clustering, in artificial and synthetic datasets. Our results show that Clus-EDA often outperforms the baseline algorithms with regard to distinct cluster validity criteria.