A L2 Discrepancy Learning Process with Applications to Outlier and Insider Detections With Large High-Dimensional Data

In this paper, a discrepancy-based framework is first presented for outlier and insider detections purpose. Given any sequence of profiles, a local discrepancy first identifies regions where the profiles are clumped or scarce then a global L2 discrepancy summarizes the overall distribution patterns of the data into one real value. A L2 discrepancy learning process is formulated to rank each profile in the sequence on the basis of optimizing the L2 discrepancy value. This L2 discrepancy learning process allows an access to many levels of information about outliers and insiders in the data. Experimental results are given to demonstrate the application of the L2 discrepancy learning process with different features data sets showing that the algorithm efficiently detects the outliers and insiders in the data.