Optimization of k harmonic means clustering with metaheuristics

OPTIMIZATION OF K HARMONIC MEANS DATACLUSTERING WITH METAHEURISTICS(Ph.D. Thesis)Alper ÜNLERGAZİ UNIVERSITYINSTITUTE OF SCIENCE AND TECHNOLOGYDecember 2006ABSTRACTData clustering analysis is the basic problem of data mining and vectorquantization. There are basically two types of clustering: Hierarchical andpartitional. Center based clustering methods are the most common ones inpartitional clustering. K-Means clustering algorithm forms the basis of thecenter based clustering. Although it is very easy to implement and understand,K-Means clustering algorithm, it suffers form two major drawbacks; Firstly itis sensitive to initial conditions and secondly it converges to a local optimum. K-Harmonic Means data clustering method considerably solves the sensitivityproblem. Since the objective function is nonconvex and there exists many localminima, converging to a local optimum is still a problem of K-Harmonic Meansdata clustering. In this study, some heuristics like tabu search, simulatedannealing and particle swarm optimization are integrated with K-HarmonicMeans method to solve the local optimum problem. Three hybrid algorithmsare developed and implemented during this PhD study. The names of thealgorithms are TABAKHO, TAVBEKHO and PARSEKHO.The hybrids developped in this study are tested on the well known datasets and inmost cases they overperformed the alternative methods.Science Code : 906.1.148Key Words :Data clustering, K-Means, Fuzzy K-Means, K-HarmonicMeans, tabu search, simulated annealing, particle swarmoptimizationPage Number :132Advisor :Prof.Dr.Zülal GÜNGÖR