The Research And Implementation Of Multiobjective K-harmonic Means Clustering Algorithm Using Swarm Intelligence

Clustering analysis is one of most important technologies in the Data Mining.The goal of Clustering analysis is to find a partition of datasets, so as to minimizeor maximize the objective function value. So the aim of clustering analysis is tofind a result so that a restricted condition is optimized. Since clustering analysis isan unsupervised learning method, it needs clustering validity indices. K-harmonicMeans (KHM) is a partition based clustering method. It uses harmonic means toevaluate the distance of a point to the clustering centroids. It uses the probabilityof clustering centroid to data points and evaluates weight dynamically in eachgeneration, thus makes the clustering softly.But there are3main weaknesses in KHM clustering algorithm. First, it issensitive to random initialization, and will converge to local optimal easily. Second,KHM optimizes single objective function. Finally, when the dataset grows, theruntime of algorithm will be too long to wait. The research work in this paperfocuses on solving these3problems using different methods.(1) In recent years, many researchers have made a lot of research to improveKHM algorithm. They used traditional evolution algorithms, such as GeneticAlgorithm, Differential Evolution, Simulated Annealing and Particle SwarmOptimization help KHM overcome local optimization. But recent studies showthat Levy flight based Cuckoo Search algorithm, proposed in2009, has even betterglobal search ability and it can be used in clustering algorithm. Thus, KHMalgorithm using Cuckoo Search based on Levy flight will overcome local optimizeand converge faster. The result of experiments show that Cuckoo Search basedKHM has better clustering results and convergence than KHM algorithm.(2) Multi-objective optimization technology is used in KHM algorithm inorder to solve KHM optimized single objective function. The MOKHMCSalgorithm is proposed, it optimizes the KHM objective function and Xie-Beni index simultaneously. Some modifications have been made in Cuckoo Search toapply to MOKHMCS. It will produce near-Pareto optimal solutions in the finalgeneration and PBMF index is used to select a final solution. The result ofexperiments shows the effectiveness of MOKHMCS algorithm and Levy flightbased Cuckoo Search helps KHM escape from local optimal and speed upconvergence.(3) The research work on parallelism of MOKHMCS algorithm is inspiredfrom cuckoos search nests in the nature and MOKHMCSP algorithm is proposed.Partition a swarm of cuckoos into many small groups. The search process betweeneach group is parallel, and serial in every group. Thus, increasing the degree ofcomputation environment will increase the speedup of algorithm. The experimentsare performed in3different computation environment, using swarm size of20and80on10different datasets. The experiment results show that, MOKHMCSPalgorithm has very good speedup on some datasets, but no good on the highdimension gene expression datasets. Some preprocessing method such asdimension reduction technology should be used when clustering. When the solingproblem increasing and creasing the degree of parallelism, the MOKHMCSPalgorithm will also has good value of speedup. It shows the expansibility ofMOKHMCSP is very good.The research work in this paper use different solution to solve the weaknessof KHM clustering algorithm. All the experiment results show the research workhas theoretical significance and good application value especially the parallelclustering algorithm MOKHMCSP.