Identification of Multivariate Outliers

Outliers, or unusually extreme values, havetraditionally been viewed as a nuisance toresearchers. Classical statistical analysis can leadto completely opposite conclusions if outliersare present or absent. Such points can, however,alert the researcher to unexpected features hiddenwithin in a data set, and lead down paths ofsurprising discovery. Outliers could even be theprimary purpose of the investigation. Credit cardfraud, electronic network intrusions, andunusual stock characteristics preceding a large move,for instance, can all be seen as outliers whosepresence is important to establish as quickly aspossible. Several methods have beenproposed to identify outliers, but many of these arenot computationally suitable for large data sets.This book presents a review of multivariate outlieridentification with particular emphasis on large datasets, and investigates a new method. The intendedaudience is statistics practitioners and dataanalysts who wish to detect outliers, as well asthose interested in the historical development of thefield. Basic familiarity with statistical concepts isassumed.