Research On Clustering Ensemble Algorithms And Their Applications

Clustering analysis is the assignment of a set of data objects into subsets so that data objectsin the same cluster are similar in some predefined sense. Clustering is a method of unsupervisedlearning. There are no predefined classes or training samples to describe the interrelationbetween data objects. Nowadays, clustering analysis has various applications in many fields,including data mining, machine learning, pattern recognition, image analysis and bioinformatics,etc. And it has been received more and more attentions. Although numerous algorithms areproposed for clustering, there are, respectively, some intrinsic limitations such as not highlyscalable, failing to discover clusters with arbitrary shape, sensitive to the order of input data, andso on.Clustering combinations was originally introduced by Strehl and Ghosh in2002. Itcombines multiple partitionings of a collection of data objects without accessing the originalfeatures. It can go beyond what is typically achieved by a single clustering algorithm in a gooddeal of respects such as robustness, novelty, stability, parallelization and scalability. Despite alarge amount of experimental results have shown that the combination clustering are better thansingle clustering methods, it still far from mature. The main research contents and contributionsof this paper are summarized as follows:(1) On the basis of having studied the fundamental techniques of clustering combinationthoroughly, a new clustering combination algorithm, Non-negative Matrix Factorization basedClustering Combination Algorithm (NMFCCA), is proposed in this paper. It not only designedfor the "hard-hard" partitionings, but also suitable for the "soft-hard" or "soft-soft" ones.Moreover, with the help of NMF, we can extract features of data objects from clusterers set andreduce the influence of noise as well. In the end, the comparison experimental results haveshown that the proposed algorithm is feasible and effective.(2) In this paper, the application of clustering combination on medical image segmentationwas introduced. As known to all, medical image segmentation is a long standing problem. Thus,it is very difficult to design a general purpose algorithm. The goal of medical imagesegmentation is to obtain a labeled image where each label corresponds to a salient image region.So it can be thought of as clustering. To improve the segmentation quality, clusteringcombination is a natural solution. It obtains a final segmentation result after feed the differentimage segmentation results to the consensus function. The experimental results on lymphoidtumor cells image and the breast tumor cells image have shown that the segmentation performance can be improved effectively.