Research On Hybrid Evolution Clustering Algorithms And Applications

With the development of society, more and more information is constantly emerg-ing, some of which will be stored and some presented as data for further analysis andmanagement. In order to understand the complex phenomena behind these data, it isvery important to classify these data or gather them into a group or a cluster. Clusteringis a highly complicated problem, thereof the concern of traditional clustering algorithmsis reduced because of its slow convergence and sensitivity to initial values. However,the evolutionary algorithm is widely considered very efective to gain a relatively opti-mal solution within a reasonable time for NP-hard problems. In the last decade, theclustering algorithms which are based on swarm intelligence have gradually replaced thetraditional clustering techniques. Evolutionary algorithm is usually a good choice incase that heuristic information cannot be applied or that the results guided by heuristicinformation are not satisfactory, because evolutionary algorithms, compared with otherglobal optimization techniques, are easier to implement and can provide adequate al-ternative feasible solution. According to the statistical learning theory, the accuracy ofclassifcation can be improved and the structural risk will be minimized by mixing someclassifcation algorithms; similarly, mixing can also be used to improve the accuracy andreduce the risk while solving clustering problems. To better improve the performanceand solution quality, hybrid evolutionary clustering algorithms have recently been pro-posed as a hot research topic. As the data of clustering problem is not labeled, thequality of the solution is generally evaluated by the validity index. The validity indexcan also be used as ftness function to guide the optimization process regarding evolutionclustering. For high-dimensional clustering problems, the interference will deviate fromthe correct direction due to the curse of dimensionality and some irrelevant informationin the attribute; thus, it is very important to research dimension reduction in clustering.This paper features the following innovative ideas and achievements:Summarized the framework of the hybrid evolution clustering algorithm and de-signed three new hybrid evolutionary clustering algorithms, namely, NichePSO,KFCMACO, and SADE. Experiments show that these new methods, compared withthose single algorithms, can reduce the structure risk of clustering under the situ-ation of prior knowledge absence. Furthermore, Experiments also show that thesenew algorithms are better than the existing hybrid evolution clustering method PSOSA.Proposed the concept and framework of hybrid validity index, that is, the hybridvalidity index of committee and hybrid validity index of penalty function. Accordingto the framework, the WCOMM and OPnealt indices are designed. Experimentsshow that these indices used as ftness function in evolution clustering in the absenceof a prior knowledge can reduce the structural risk.Researched the problem of dimension reduction in evolution clustering and designedtwo algorithms, namely, LDAPSO and MDSACO. Experiments show that the di-mensionality reduction can improve the accuracy of the high-dimensional clusteringwhile using these algorithms. From experiments, it also shows that the two algo-rithms are better than the existing method MDACO.Used default logical reasoning to do the non-monotonic generalization on clusteringresults and thus avoided the loss of efective information while using Occam razor.Expanded the applications of clustering. Hybrid evolutionary clustering is success-fully used in the online reputation rating evaluation and the partition of Sedimentarymicrofacies in reservoir description.