Evolutionary Clustering Using Multi-prototype Representation and Connectivity Criterion

Abstract

An automatic clustering approach based on differential evolution (DE) algorithm is presented. A clustering solution is represented by a new multi-prototype encoding scheme comprised of three parts: activation thresholds (binary values), cluster centroids (real values), and cluster labels (integer values). In addition, to measure the fitness of potential clustering solutions, an objective function based on a connectivity criterion is used. The performance of the proposed approach is compared with a DE-based automatic clustering technique as well as three conventional clustering algorithms (K-means, Ward, and DBSCAN). Several synthetic and real-life data sets having arbitrary-shaped clusters are considered. The experimental results indicate that the proposed approach outperforms its counterparts because it is capable to discover the actual number of clusters and the appropriate partitioning.