Multi-Scale Classification Method

As an important branch of data mining,the essence of classification mining is through learned model,training by the known data,to forecast the label of unknown data.Multi-scale classification mining is a typical interdisciplinary topic,aiming at combing multi-scale science with classification technology,analyzing the character of data in multidimensional,obtaining classification models at different levels,as well as learning more comprehensive information.At present,the research of multi-scale classification has made some achievements,which can effectively improve the classification performance,but it is mostly limited to image and spatial data.Previous studies have shown that multi-scale research on general data sets has also achieved initial results,such as multi-scale clustering,multi-scale association rules,but few studies are conducted in the field of classification.In order to solve the above problems,this thesis tries to study the universal multi-scale classification mining method,which can not only expand the scope of multi-scale application,but also improve the classification efficiency.From the perspective of spatial data estimation,combined the hierarchical theory and scale characteristics,and based on the discretization method of probability density estimation,this thesis studies the transformation method of multi-scale classification mining under general data sets.Classification mining is carried out according to the multi-scale characteristics of data.Based on the theory of non-local mean and double cube interpolation,using Q statistics and inconsistent measurement to operate,this thesis proposes the upscaling algorithm of multi-scale classification and downscaling algorithm of multi-scale classification.The main works are stated as follows:(1)Research on the theoretical basis of multi-scale classification.Firstly,based on the characteristics of classification mining and the discretization method,this thesis studies the method to select division scale and determine the number of the scale layers,according to the scale characteristics and the equivalent partition model;Secondly,it analyzes the range scale and granularity scale to construct the multi-scale data set,referring to the score of the partition point which is calculated by the attribute value of the characterization scale;Thirdly,it combines three-way decision with multi-scale to select the benchmark scale by using the multi-scale decision table;Finally,the thesis proposes the essence and architecture of multi-scale classification,which provides the theoretical basis and method for the subsequent research of multi-scale classification mining.(2)The proposal of multi-scale classification algorithm.UAMSC(Upscaling Algorithm of Multi-Scale Classification)and DAMSC(Downscaling Algorithm of Multi-Scale Classification)are proposed based on the theory of spatial data estimation and scale conversion.In the upscaling,based on multi-scale self-similarity and smoothing detail information,this thesis adopts the idea of non-local mean weighted filtering and uses Q statistics to realize knowledge derivation from small scale data set to large scale data set.In the downscaling,based on the disagreement measure,the thesis uses the theory of cubic convolution interpolation to increase the detail information,specify the mining knowledge,weighted to obtain the target scale information,and realize the knowledge derivation from the large scale data set to the small scale data set.(3)Verification experiments on the framework and algorithm of multi-scale classification.The experiments are carried out on four UCI common benchmark data sets and one real data of H province to verify the feasibility and effectiveness of the algorithms.The results show that UAMSC and DAMSC algorithms are feasible and effective,with high accuracy and low time complexity.They are not only better than LIBSVM and other benchmark algorithms in efficiency,but also better than MSCSUA and MSCSDA in performance.