Research On Fast Attribute Selection Approach In Complex Granulation Environment

With the rapid development of information technology,the Internet has become more and more closely connected with other industries,and vast amount of data are continuously generated in real applications,while the data collected in practice increasingly presents a variety of complex characteristics such as large data volume,high feature dimension,low data value density,etc.This will make the cost of data storage too high and the acquiring of the potential value of the data more difficult.Therefore,how to mine large-scale and complex data for acquiring knowledge is a very meaningful research topic.Attribute reduction,as an effective data reduction and knowledge discovery technology with clear semantic interpretation,has been widely concerned and successfully applied into many fields.As a widely accepted intelligent processing paradigm,Granular Computing simulates the granulation and hierarchical cognitive mechanism of human thinking and solving complex problems,and through granulating the formation in data and performing the computing of granules,a deeper interpretation of data can be achieved.Therefore,data mining and knowledge acquisition,coupled with the technology of Granular Computing,can provide effective solutions for the analysis and processing of large and complex data.However,based on the practical requirements,the problem is frequently analyzed and solved from the perspectives of multi-level,multi-view and multi-granularity,and data modeling and rough analysis in such environment may face challenges such as high computational complexity,poor time performance,etc.In view of this,how to conduct fast attribute selection in complex granulation environment is discussed and studied in this dissertation thoroughly,and the corresponding algorithm frameworks are developed and constructed on the basis of this study.Specifically,the main researches are show as following.(1)The Gaussian kernel function is introduced into the constructing of fuzzy relation for the information granulation of continuous data,and multiple kernel parameters are used for observing the generalization performance of the selected subsets of attributes from different granulation scales.To improve the time efficiency of dimension reduction,a fast attribute selection approach is designed from the viewpoint of parameter.Experimental results over several models demonstrate that the acceleration method can generate the subset of attributes with comparable generalization performance at lower time consumption.(2)The fuzzy relation can be used for characterizing the similarity between sample instances,and the supervised information has not been considered in its construction process,then the information in granule may be loss if supervised strategy is directly introduced for fusing and improving the discrimination ability and learning ability of the granulation result.Consequently,the scope of information granulation is expanded,and both neighborhood information granulation and pseudo-label information granulation strategies are used for data modeling and rough analysis.The corresponding acceleration approach is further design,which can effectively reduce the time consumption of attribute selection process.(3)From the viewpoint of granularity,the existing attribute reduction researches can be divided into two main categories: single granularity based attribute reduction and multi-granularity based attribute reduction,and it is further clarified that the difference between them lies in the constructing of constraint.It is also pointed out that most of works related to the acceleration approach of attribute reduction have only considered from the perspective of sample,while few researches have been reported from the view of attribute.Therefore,from the perspective of attribute,a strategy for fast attribute selection is designed based on the relationship between attributes.Finally,two types of efficient attribute selection algorithm frameworks are presented in terms of the dissimilarity and similarity between attributes,which can effectively improve the time efficiency of knowledge acquisition.Moreover,combined with the existing method,the proposed algorithm frameworks can achieve better time performance.