Research On Outlier Detection Methods Based On Neighborhood Rough Measure

Outlier analysis is an important research branch in data mining technology,and its purpose is to find some minority data objects that are inconsistent with general behaviors or patterns.In most data mining research,outliers are treated as noise and discarded.However,in the application research of fraud detection,medical detection,intrusion detection,public safety,and image processing,the emergence of such minority events may be more interesting than those that normally occur.Granular computing is an intelligent computing paradigm for information processing,which applies the information granulation mechanism to the reasoning process of humans to solve practical problems.Neighborhood rough set theory is an important type of granular computing model,which has been proved to be an effective tool for feature selection,rule extraction and knowledge discovery in numerical or mixed attribute data analysis.Recently,outlier detection based on neighborhood rough set theory has been proposed.However,these methods do not involve the rough approximation accuracy of the neighborhood and the rough distance of the neighborhood for outlier detection.The rough approximation accuracy of the neighborhood and the rough distance of the neighborhood are effective uncertainty measurement methods in the neighborhood rough set,which can be used to build an outlier detection model for mixed attribute data.To this end,for the problem of outlier detection of mixed attributes,this topic will focus on the research of outlier detection methods based on neighborhood rough metric.The main research contents are as follows.Firstly,the neighborhood information system is constructed with the optimal heterogeneous neighborhood relation measurement and the statistical neighborhood radius.Then,aiming at the shortcomings of the outlier detection method based on rough overlap metric,an outlier detection method based on the rough overlap metric of neighborhood is proposed.This method defines two overlapping distance measurement methods for mixed attribute data in the neighborhood rough set and defines corresponding outlier factors to detect outliers of mixed attribute data.The effectiveness of the proposed new method is verified by theoretical analysis and comparative experiments with UCI data.Furthermore,based on the concept of neighborhood approximation accuracy,an outlier detection method based on neighborhood approximation accuracy is proposed.This method uses the neighborhood approximate precision outlier factor to characterize the outlier degree of the data object.UCI data comparison experiment results show that this method can be suitable for outlier detection of data with various attribute types.