Multi-scale Clustering Method Based On Coupled Metric Similarity

With the development of multi-scale data mining,the application of multi-scale analysis in the fields of remote sensing image recognition,disease detection,fault diagnosis,and cluster analysis is more and more mature.However,the existing multi-scale data mining is mainly for quantitative analysis and prediction of numerical data sets.There is little research on data sets,especially for the intrinsic similarity measurement of not independent and identically distributed categorical data sets.This dissertation combines multi-scale clustering theory and similarity measurement methods,and starts with multi-scale clustering tasks to construct a clustering model of categorical multi-scale data sets.Using unsupervised coupled metric similarity method,a benchmark scale clustering algorithm is proposed for categorical data sets that are not independent and identically distributed.Based on the idea of agglomerative hierarchical clustering and Lanczos interpolation theory,a scale transformation model is constructed to effectively reduce the scale effect in multi-scale clustering data mining.In this thesis,the clustering analysis is carried out based on not independent and identically distributed categorical multi-scale data sets.The main work of this thesis includes the following aspects:(1)Research on the multi-scale clustering theory of categorical data.Based on the coupled metric similarity method,a multi-scale clustering method for categorical data is proposed by considering the intra-attribute similarity and intre-attribute similarity.The multi-scale clustering method is extended to the field of categorical data,and is perfected to provides a theoretical basis for subsequent multi-scale cluster analysis of the categorical data.(2)Construction of the multi-scale clustering algorithm framework based on coupled metric similarity method.Based on coupled metric similarity,a multi-scale clustering mining method for categorical data sets is proposed.Firstly,not independent and identically distributed data set is preprocessed.Then,the most suitable benchmark scale is selected based on the scale division and benchmark scale selection methods.Finally,according to the coupled metric similarity theory,the multi-scale clustering architecture is constructed.(3)Proposal of multi-scale clustering algorithm.Combining the method of data set partition and benchmark scale selection methods,a benchmark scale clustering method is proposed based on coupled metric similarity method;based on the idea of agglomerative hierarchical clustering,a multiscale clustering up-scaling algorithm is proposed;the essence of Lanczos interpolation is analyzed,and the contribution rate of the known sample from different aspects is considered;combined with the idea of scale up based on cubic convolution,a scale down algorithm for multi-scale clustering is proposed.(4)Verification and analysis experiments on the multi-scale clustering algorithm based on coupled metric similarity method.The multi-scale clustering algorithms is analyzed using UCI and Kaggle public data sets and one real total population data of H province.The experiments are carried out by comparison with the CMS,IOF,HM and other similarity metric methods combined with spectral clustering.The results show that the indicators of the benchmark scale clustering algorithm based on the coupled metric similarity method,the up-scaling algorithm based on the single chain,and the down-scaling algorithm based on Lanczos interpolation in the NMI,MSE,and F-score have different degrees of improvement,and it has a shorter running time.Experiments prove that the proposed multi-scale clustering algorithm based on coupled metric similarity is effective and feasible.