The Research On Clustering Algorithm Based On Particle Swarm Optimization And Rough Set Theory

Data mining is one of the most forward lines of database and information managementarea which excavate the potentially valuable information from massive datas.Cluster analysisas an important part of data mining,has been used in pattern recognition,imageprocessing,data compression ,marketing and others fields.This paper discussed the K-medoids clustering algorithm、particle swarm algorithm, thekernel function and rough set,The main research of this paper as follows:(1) For the disadvantages that sensitivity to centers initialization，lower clusteringaccuracy and slow convergent speed of K-medoids algorithm, a novel K-medoids algorithmbased on density initialization, density of iterative search strategy and optimization criterionfunction is proposed. Experimental results show that the algorithm can take advantage ofdensity initialization thought, and reduce the scope of the centers to reduce the convergencetime clustering,weighted criterion function to ensure the clustering efficiency furtherly.(2)After analysis the disadvantage of local maximum of K-medoids algorithm, a novelK-medoids clustering based on particle swarm optimization is proposed.By looking for thecontraction between particle swarm optimization (PSO) and K-medodis, using the globaloptimization to prevent the algorithm into local optimum. Experiments show that thealgorithm has higher accuracy, smaller time complexity and more stable overall performance.(3)A rough kernel clustering algorithm based on PSO is proposed,which combine withthe K-medoids,rough set, kernel function and PSO effectively,and overcome the shortcomingof the K-medoids can not handle datas of non-linear and boundary.Firstly,map samples intohigh-dimensional space by Mercer kernal,so that the samples are translated to be linearlyseparable.Then,deal with bundery objects with the idea of rough set and make samplesattributes weighting processing with ReliefF.Finally,adopt the pso to prevent the algorithmfrom trapping into local optimum.Experiments show that the algorithm has good clusteringresult on the unsub-liner and high-dimensional datas, and prove the algorithm’s efficiencyand accuracy.