Clustering Validity Analysis And Its Application In Electrical Tomography

Cluster analysis is a main task of exploratory data mining and an important content of unsupervised learning.Cluster analysis or clustering aims to group similar patterns into a class and separate dissimilar patterns for further analysis and decision making.This paper mainly includes the data reduction,design of cluster validity index and application of clustering in electrical tomography.1.An efficient data reduction method based on grid-based bisecting algorithmA new efficient measure was developed to reduce a given dataset and to uncover the major features by multiplying the defined absolute density with the defined local density of any data.The absolute density is similar to the density measure in many other clustering algorithms,but the proposed method can estimate the absolute density in a faster unsupervised manner.The local density overcomes the problem that the absolute density can not extract the local clustering structure in the existing algorithms.These two kinds of densities were estimated with a fast grid-based bisecting method.A group of feature-different synthetic datasets and benchmark datasets were used to test its performance on the clustering accuracy,runtime and separability among clusters of feature reduction and sample reduction.The results strongly proved that the proposed method could fast reduce a dataset and identify the most important key features.Additionally,it also can effectively determine the optimal number of clusters by suppressing the noisy data and enhancing the separation among clusters.2.A new clustering validity index based on the dual centerBased on two most commonly used partitional clustering algorithms,c-means and fuzzy c-means,and their variants,a new measure which is called as dual center,was developed to represent the separation among clusters.The new meaure can efficiently represent the separation among various clusters.According to the defined measure,a new validity index is proposed for evaluating the clustering performance of partitional algorithms.Two groups of benchmark datasets with different characteristics were used to validate the effectiveness of the proposed validity index.Experimental results provide evidence that the proposed valdity index outperforms some existing representative validity indexes in the two groups of typical and representive datasets.3.An unsupervised evaluation method of image quality based on fast fuzzy clustering algorithmIn practice,the distribution of the measured object is unknown and an unsupervised image quality evaluation method is needed,so,based on fast fuzzy clustering algorithm,an unsupervised image quality evaluation method for electrical tomography was proposed.By fast fuzzy clustering algorithm,the reconstructed images of the Tikhonov regularization algorithm and conjugate gradient algorithm were segmented to the target and the background,respectively.Find the representative gray values of target and background,and calculate the deviation between the target and the background pixels and their representative gray values,respectively.The regularization parameter or the number of iterations was determined by the minimum value of the deviation curve,and the image quality was evaluated.The experimental results show that the proposed method is consistent with the commonly used image relative error method and correlation coefficient method and it is unsupervised,fast,and robust.