An Algorithm for Clustering of Outliers Based on Key Attribute Subspace

It is an important part of data mining to discover and analyze outlying observations. Outliers may contain crucial information, and so detecting them is much more significant than detecting general patterns in some applications which include, for instance, credit card fraud in finance, calling fraud in telecommunication, intrusion in network, disease diagnosis, etc. Existing outlier mining algorithms focus on detecting and identifying outliers, but studies of outliers include both mining outliers and analyzing why they are exceptional. The research on explaining and analyzing outliers slightly lags behind outlier mining technology now. It is inevitable that analyzing outliers to the full needs a great deal of knowledge from object task fields. However, some further discoveries of outliers may be obtained from studies of distributing characteristics of dataset in attribute space. By analyzing the origin and feature of outliers and using the theory of rough set, a concept of outlying partition similarity is defined and then an algorithm for clustering outliers based on key attribute subspace (COKAS) is proposed. The approach can provide the extended knowledge of identified outliers and improve the understanding of the whole data set. Experimental results of real multi-dimension data set show that this algorithm is scalable and efficient.