Fast Clustering Convergence by Genetic Algorithm

The work described in this paper aims to experiment with unsupervised clustering using a developed tool based on Franti’s genetic algorithm. The tool is characterized by a seeding meta-method exploited to initialize the population solutions. In addition, the evolution of the population is strictly based on an elitist approach which at each step selects pairs of solutions to be crossed among the best solutions of the current population generation. Such solutions are then crossed/merged into a new solution by a fast pairwise-nearest-neighbor (PNN) smoothing technique which, after a local optimization by Lloyd’s K-Means, enters to compose the next generation of the population. The tool is prototyped in parallel Java ensuring fast convergence in practical cases. An in-depth experimental work confirms high-quality clustering and very good execution performances on both synthetic and real-world datasets.