Outlier Detection in Complex Structured Event Streams

Outlier detection methods are now used extensively, particularly in systems for detecting internal intrusions, in medicine, and in systems for detecting extremism in public political discussions on forums and social media. The aim of this work is to consider a fuzzy method of detecting outliers, based on elliptic clustering in the higher-dimensional space of attributes and using the Mahalanobis metrics for calculating the distances between objects and the center of a cluster. A procedure developed by the authors is used to find the optimum values of metaparameters of this algorithm. The classification of both individual events and complete sessions of user activity is considered, using an algorithm based on Welch’s t -statistics. The proposed procedures display a high quality of operation in solving two important problems of the stream analysis of complex data structures: the authentication of users by keystroke dynamics, and detecting extremist information in web text messages.