Research On Incremental Mechanisms And Robust Algorithms Of Attribute Reduction

With the development of information technology and the popularization of intelligent terminals,human society has gradually stepped into the era of artificial intelligence.The way of data generation and collection has undergone great changes,presenting the characteristics of large-scale,dynamic updating and containing noise.It has become the focus of attention in the filed of mechine learning to design efficient dimensionality reduction algorithoms with robustness.Rough-set based attribute reduction removes redundant condition attributes from the data sets to achieve data dimensionality reduction by preseving the inconsistency between condition attributes and decision labels.Due to the increasing scale and dynamic updating of data sets,the current attribute reduction algorithms face the following challenges:one is how to realize attribut reduction for large-scale dynamic data sets under the existing hardware environment;the other is how to improve the robustness of the positive regioin preserved reduct in view of data disturbance.This dissertation studies the incremental mechanism of attribute reduction,and studies the theory and algorithm of robust attribute reduction based on machine learning regularization method.The detailed research contents and results are presented as follows:(1)The incremental mechanism of attribute reduction with classical rough sets is studied when samples and attributes are added to dynamic data simultaneously.Through discernibility relation,the judgment mechanism is designed to judge whether the adding new samples and attributes are necessary for updating the current reduct,and the mutual restriction relationship between samples and attribute is revealed in dynamic data at the mean time.Unnecessary samples and attributes are stored in data instead of being filtered.Necessary samples and attributes are selected to update the reduct through incrementally calculating the discernibility relation.Based on the unified incremental mechanism,an incremental algorithm of classical rough sets is designed for attribute reduction with samples and attributes increasing at the same time.Experiments verify the efficiency and effectiveness of the algorithm in runtime and accuracy.(2)The incremental mechanism of attribute reduction with fuzzy rough sets is studied when samples and attributes are added to dynamic data simultaneously.The relative discernibility relation is proposed to depict the samples and attributes of fuzzy rough set in a unified way.It provides a theoretical basis for the incremental mechanism analysis of the simultaneous increase of samples and attributes.By analysis the variation of the relative discernibility relationship caused by the increase of samples and attribtues,the incremental mechanism of attributes and samples increase are well intergrated to form an incremental mechanism of simultaneous increase of samples and attributes.When new samples and attributes are added continuously,a judgement mechanism is designed to judge whether it is necessary to update the current reduct,so as to delete the redundant attributes and add the necessary attribute.Based on the unified incremental process,an incremental algorithm is designed for dynamic data sets.Numerical experiments show that the algorithm can effectively deal with the incremental attribute reduction with attributes and samples added simultaneously.(3)The theory and method of robust attribute reduction are studied.The positive region preserved reduct is not robust,for the rough sets are sensitive to data noise.Inspired by the regularization method in machine learning,the robust attribute reduction theories and methods of classical rough sets and fuzzy rough sets are studied respectively according to different data types.The regularization strategy balance the empirical error of the decision rules and the granularity of the information table,which improves generalization ability of extracting decision rules at the expense of the decision rules with lower support in the information table.Based on the optimization strategy of regularized loss function,a robust attribute reduction algorithm is designed to further delete the attributes with weak discerning ability in positive region preseved reduct,thus improving the generalization ability of the decision rules extracted from the decision sysytem.The robust attribute reduction algorithm is applied to the selection of key features of coal consumption in power plants,and the selected key features are used as input variables to establish the experimental model.Experimental results show that the robust attribute reduction algorithm can effectively delete the attributes with weak discerning ability and improve the generalization ability of extraction rules.