A Multi-objective Clustering Ensemble Method Based On The Minimum Spanning Trees And Its Application

As one of the main composition of information technology development, data mining is a process from which we can reveal hidden, unknown but have potential value information. As an important branch of data mining techniques, clustering analysis gets more and more attention. A lot of clustering algorithms have been proposed by experts and scholars and widely used in real life. Nowadays, with the increasing number of data size, the high-dimensional data clustering algorithms become the difficult point of academic research. When dealing with high dimensional data sets, traditional clustering algorithms often lead to “dimension disaster”. Therefore, ensemble technology in machine learning is introduced into the clustering algorithms by experts, called cluster ensembles. Thus, the accuracy and stability of clustering algorithm are greatly improved when dealing with high dimensional data sets.Based on the background above, a multi-objective clustering ensemble method based on the minimum spanning trees is proposed by this paper, and is applied in a large number of gene expression data sets, UCI data sets and SAR image segmentation. The specific works of this paper are concluded as follows:1. A multi-objective cluster ensemble method based on the minimum spanning tree is proposed, and the framework of the MOCLE is adopted by this algorithm. Firstly, a new crossover operator which is based on the minimum spanning tree is applied. It overcomes the defect of the original crossover operator which produces illegal solution in the process of evolution. Secondly, a new operator is proposed, which can get new cluster population by searching the subspace of the data sets using different traditional cluster algorithms. Thirdly, a new objective function which represents the similarity between the individual to be evaluated with the others is added. Finally, a new selection strategy is added before the process of crossover, which can select the best quality part of the based clustering results. In the experiments, a large number of gene expression data sets and UCI data sets are employed to test the method. Our algorithm is compared with several similar algorithms, and the experimental results prove the new proposed algorithm is greater than others on getting good solutions.2. An improved multi-objective cluster ensemble method based on the minimum spanning tree is applied in image segmentation. The method is based on the improved algorithm called MOCMST. Firstly, gray level co-occurrence matrix method and the undecimated wavelet decomposition method are respectively used for the extract of texture feature vectors and wavelet vectors. Secondly, the watershed algorithm based on the gradient is used to the beginning of segmentation, so as to translate the pixels into super pixels. Thirdly, the method searches the gathering condition of the point by multi-objective cluster ensemble method based on the minimum spanning tree, determining which region the point belongs to and giving the pixels corresponding labels so as to realize the image segmentation by the clustering algorithm. In the experiments, the method has been tested on four SAR images. Compared with other two methods, the proposed method has a good performance for SAR image segmentation.