Attribute Reduction Acceleration Model For Multi-granularity Problems

With the rapid development of computer technology,the Internet,human production and life are increasingly integrated.To describe one field as completely as possible,it is necessary to investigate and record many samples and attributes of each sample,then large-scale,mixed-structure and high-dimensional mass data will be generated.As a key step of data preprocessing,attribute reduction has been effectively applied to many fields such as rough set,granular computing,pattern recognition,data mining and so on.Attribute reduction can help us select the smallest attribute subset without relevant and redundant attributes,and the subset satisfies the given conditions.However,with the advent of the data age,data is no longer presented as a single form,and there may be some limitations while attribute reduction is used to processing data.To overcome these limitations,in this dissertation,the cognitive mechanism of the human brain will be learnt to deal with complex problems comprehensively.On the basis of the definition of traditional attribute reduction,the high-dimensional complex data is processed from multi-granularity,multi-view and multi-level.Analyze the definition of attribute reduction and the strategy of compting reduct,then construct attribute reduction which is suitable for multi-granularity,multi-view and multi-level,and explore the method of computing such reduct.Furthermore,optimize the method,and improve the time efficiency of computing multi-granularity reduct.Specifically,the research content and innovative methods of this dissertation mainly cover the following three points.Firstly,the parameterized multi-granularity attribute reduction and its acceleration strategy are proposed.Attribute reduction is focused on the single granularity perspective,and computing reducts on the basis of different granularities one by one.Then the variation tendency of generalization performance provided by reducts can not be observed intuitively,and the method of computing reducts over multiple granularities is time-consuming.In view of this,multi-granularity is constructed by using multiple different parameters,and parameterized multi-granularity attribute reduction is proposed.Furthermore,considering two variations of parameterized granularity: from fine to coarse and from coarse to fine,two acceleration strategies of computing parameterized multi-granularity reducts are designed.These methods compute reduct over the current granularity on the basis of reduct over the previous granularity,it follows that the searching space of attributes will be compressed and the time consumption will be reduced.Secondly,the supervised neighborhood based multi-granularity attribute reduction is proposed.In the above parameterized multi-granularity attribute reduction,if the parameterized multi-granularity is constructed in the framework of neighborhood relation,then the radius will be used to determine whether samples are similar.However,by using this method,two samples with different labels may fall into the same neighborhood and are regarded as indistinguishable.In view of this,when judging whether samples are similar,the true label of sample is taken into consideration,and the supervised neighborhood relation is proposed,then intra-class radius and inter-class radius are used to determine whether samples are similar,it follows that the defect that samples with different label fall into the same neighborhood will be alleviated effectively.Furthermore,the acceleration strategy for computing parameterized multi-granularity reduct is introduced into the process of computing supervised neighborhood based multi-granularity reduct,and the time consumption of computing such reduct will be reduced effectively.Finally,the sample based multi-granularity attribute reduction and its acceleration strategy are proposed.Most of existing attribute reduction is based on a single and fixed sample structure,and does not consider the universal of the reduct when granularity changes slightly due to sample perturbation.In view of this,the multi-granularity structure is constructed by using different sample structures,and the definition of sample based multi-granularity attribute reduction is proposed.Considering the guidance of reduct over all samples on reducts over other granularities,an acceleration method for computing sample based multi-granularity reducts is designed.This method computs the reducts over different granularities on the basis of the reduct over all available samples,which effectively improves the time efficiency of computing sample based multi-granularity reducts and improves the stability performances related to reducts.