Research On Quick Attribute Reduction From Local Perspective

With the advent of the third generation of Artificial Intelligence,various technologies have also developed rapidly such as Data Mining,Pattern Recognition,and Cloud Computing.At the current stage of data-driven artificial intelligence,aiming at practical application,finding the important information has become the main driving force in the development of advent science and technology.As an effective tool for data mining,attribute reduction is dedicated to removing the redundant attributes with a given constraint.And attribute reduction has made some contributions to reducing data storage space and saving time of data analysis.At the same time,the classification performance related to the reduct is also the focus of many scholars.In this paper,some investigations are conducted on how to speed up the process of computing reduct without reducing the classification performance.In details,the motivations and contributions are listed as follows:(1)Construct the multi-criterion attribute reduction based on fitness function.A single criterion corresponds to a single constraint,and using a single constraint to calculate the reduct may only make the reduct prossess a certain ability.It is difficult for the reduct be suitable for other multiple requirements in the complex environment.To fill such a gap,a multi-criterion attribute reduction based on fitness function is constructed.Primarily,heuristic searching strategy based on greedy strategies are used in this paper,and the heuristic information is constructed by the fitness functions which measures the relationship between condition attributes and decision attributes.Secondly,different from the traditional naive algorithm,the multi-criterion attribute reduction combines multiple fitness functions to define the constraints and then select condition attributes.Finally,to reduce the elapsed time of computing reduct,an improved version of the cluster-based sample selection algorithm is employed to compress the sample searching space.The experimental results show that:(a)compressing the sample searching space does can reduce the elapsed time;(b)compared with the single criterion reduct,the multi-criterion reduct can improve the classification accuracy with keeping the original single criterion performance.(2)An integrated selector is constructed from local view and an attribute reduction over consistent samples is proposed.Multi-criterion attribute reduction based on fitness function is an implementation strategy of the integrated ideas.However,scholars often consider the performance of decision systems as a whole but ignore that different decision classes may require different key attributes.Based onthis point,an integration strategy from local view is constructed.It is hoped that the expected reduct can balance the needs of different decision classes as much as possible.When it comes to the problem of consistent samples,the motivation and process of the consistency principle are shown in this paper,and the new decision system are employed for computing the new reduct.(3)The muti-granularity attribute reduction fretwork is constructed and the contraints under some granularities are fused.Recently,muti-granularity is an interesting topic,and different information can be obtained from different granularities or different levels.However,the existing algorithms often focus on the attribute reduction under a single granularity,which are reconsidered from the multi-granularity perspective.To construct multi-granularity attribute reduction,two main problems need to be solved are shown:(a)build a multi-granularity framework;(b)define and design multi-granularity attribute reduction.The former one is constructed by using the characteristics of the neighborhood rough set,as the neighborhood relationship constructed in the neighborhood rough set is closely related to the parameter radius,and a series of granulation results of different scales and different layers can be constructed according to different radii.In other words,multi-granularity framework can be constructed with using a series radii.The latter one focuses on the design of multi-granularity constraints and the selection of suitable candidate attributes.The multi-granularity constraints are the fusion of constraints corresponding to multiple granularities in this paper.The integration strategy from local view will be used for selecting attributes.Experimental results show that,compared with single-granularity reduct,multi-granularity reduct can save a lot of elapsed time without reducing the classification performance.Finally,the full text is summarized and the research direction is pointed out.