An Exceptional Reduction Algorithm for Outliers Analyzing in High-Dimension Space

Mining and analyzing for outliers is of great importance in many applications, including network invasion control, credit card and telecom fraud detection, etc. Existing outlier mining algorithms are focused on detecting outliers and lack valid approach for explaining and analyzing why they are exceptional. In order to describe exceptional features of high-dimension dataset in quantificational detail, the concepts of key attribute subspace of outliers and exceptional contribution degree of an attribute is defined in the paper. Furthermore, we present an idea of exceptional partition based on the theory of rough set. This leads to some efficient methods for outliers explaining and analyzing, in which an exceptional reduction algorithm (ERDA) that we proposed is mainly discussed in this paper. The ERDA offers a clever approach to identifying the origination of detected outliers and can help to improve one´s understanding of whole data set. The results from a study on its complexity and experiments on real world data sets show that the proposed algorithm is scalable and efficient