Research On Heuristic Attribute Reduction Algorithm For Neighbourhood Rough Set

The neighborhood rough set theory can be directly used to process numerical,symbolic and mixed attribute data,and has been widely used in artificial intelligence,pattern recognition and data mining.Attribute reduction is a specific application of the neighborhood rough set theory.To simplify dataset,redundant dataset attributes can be deleted under the condition that classification accuracy of original dataset being kept.By attribute reduction,not only the dataset space occupation can be compressed,but also decision errors is directly decreased.Based on the neighborhood rough set theory,some recently proposed attribute reduction algorithms have been deeply researched.The main work of this thesis is as follows:(1)Basic properties of neighborhood relation matrix have been discussed after set operations of neighborhood relationship being transformed into matrix operations,under neighborhood relationship's neighborhood relation matrix.Combined with sorting,symmetry of neighborhood relation matrix and neighborhood search idea,single attribute neighborhood relation matrix(SANRM)algorithm is proposed,which is an effective improvement version of traditional single attribute equivalence relation matrix(SAERM)algorithm.(2)Aiming at the attribute reduction problem of neighborhood decision information system,combined with sample neighborhood information granularity and its decision distribution,a new measure that can be used to effectively reflect relationships between conditional attribute subsets and decision attribute,which is constructed to solve the problem that the minimum decision criterion around the neighborhood decision error rate not being able to reflect the sample neighborhood information granularity for uniform distribution of decision classes.With the granulation monotonicity being proved,neighbor relation matrix-based attribute reduction(NRMAR)algorithm is constructed.UCI dataset based experimental analysis shows that NRMAR algorithm can be used to select attributes with good dataset classification capabilities effectively.(3)Based on consistency hypothesis of classification tasks,an adaptive class standard deviation neighborhood radius value calculating method is proposed and applied to NRMAR algorithm.UCI dataset experimental shows that the method of adaptive class standard deviation neighborhood radius value calculating method is more reasonable than fixed neighborhood radius calculating method.It can be efficiently used to obtain smaller reduction attribute subset.Therefore,value obtaining of neighborhood radius is no longer arbitrarily according to subjective experience when adaptively acquire according to dataset distribution characteristics.So it is a more general and effective method to decide neighborhood radius.