Research Of Attribute Reduction Based On Object Similarity

With the increasing maturity of AI technology and the expansion of application fields,data mining,intelligent decision-making system,knowledge engineering,distributed artificial intelligence and other fields have higher efficiency in data acquisition and data storage.However,it also makes the data types more complex and date scale rise,forming many complex types of data,such as symbolic type,incomplete type,interval type etc.Classical rough set theory is an effective tool proposed by Polish mathematician Z.Pawlak in 1982 which could describe the degree of inconsistency of classified data.It has been successfully applied in data mining,machine learning and intelligent decision-making.Attribute reduction,also known as feature selection,is one of the important research contents of rough set theory.Its purpose is to retain the minimum set of attributes that keep the classification ability of information systems unchanged,and delete redundant attributes.The concept of classical rough set model is clear,the equivalent object is represented by "1",otherwise it is indicated by "0".One of its limitations is that it can only deal with symbolic data.In view of considering that complex data types exist widely in the actual environment,this paper describes the uncertainty of interval-value data and numerical data respectively through compatibility relation and similarity relation and corresponding attribute reduction algorithm is proposed.The main tasks are as follows:1)The concept of maximum confidence is introduced in inconsistent interval-valued decision system to construct discernibility matrix whose the maximum distribution is unchangeable.We proposed maximum distributed reduction algorithm based on discernibility and analyze the relationship between maximum distribution reduction in inconsistent interval-valued decision system and positive domain and distribution reduction algorithms.Finally,experiments are carried out by using UCI standard data sets and the results of it shows the effectiveness of the algorithm.2)Focusing on some special classes of decision attributes in interval-valued decision system,the concept of local reduction in this paper.Also we give the local reduction is introduced algorithm based on discernibility matrix.Moreover,the structure of the global reduction in interval-valued decision system is depicted by the concept of the local reduction and the relationship between the local reduction and global reduction is also analyzed.Finally,the related experiments are carried out to verify the feasibility and effectiveness of it.3)Aiming at the problem of noise in the similarity relation of fuzzy rough sets,the radius of similarity relation is introduced on the basis of classical fuzzy similarity relation.The generalized decision-keeping reduction method based on discernibility matrix is designed,but its algorithm complexity is high.Therefore,two methods based on forward greed and backward greed are proposed by using similarity degree between attribute sets.A heuristic generalized decision-making preservation reduction method is proposed,and the correctness of the three algorithms is proved by experiments.Work 1)and 2)describe the similarity between interval-valued data objects by using compatibility relations;Work 3)describe the similarity between numerical data objects by using similarity relations.