B1604
Title: Multivariate outlier detection with ICS and application to statistical quality control for autocorrelated data
Authors: Stefanos Voutsinas - Athens University of Economics and Business (Greece)
Ioulia Papageorgiou - Athens University of Economics and Business (Greece) [presenting]
Abstract: Detection of special structures in the data, e.g. outliers, is an issue of high priority in Statistical Quality Control (SQC) because it can be an indicator of an out-of-control production line. The early detection of such measurements is essential and this problem is challenging when the population of interest is multivariate. Multivariate data on the other hand is more often the case than the exception in nowadays applications. Another parameter we consider for the SQC application is the autocorrelation among observations which is again a realistic scenario in this field. Most of the existing methodologies for detecting outliers fail to reveal those measurements for both multivariate case and autocorrelation. The aim is to examine the use of Invariant Coordinate Selection (ICS) in Statistical Quality Control and especially in detecting extreme measurements in case of correlated multivariate data. The experiments include various choices of scatter pairs for the use of ICS, the degree of correlation and the mechanism of generating the outliers. The performance of ICS method for detecting outliers in SQC is compared with the Mahalanobis distance and $T^2$ Hotelling chart plot. The comparison and evaluation are based on (i) the percentages of correct True Positive (TP) detection of outliers and false detection, False Positives (FP), events.
