Robust kurtosis projection for multivariate outlier labeling

Abstract

Outlier labeling can be considered as an early procedure to get the information of `suspects'. This paper introducesrobust kurtosis projection algorithm for multivariate outlier labeling of data set with moderate, high and very high percentage outlier. The algorithm works in two stages. In the first stage, we propose a projection approach to findthe orthonormal set of all vectors that maximize the kurtosis of the projected standardized data. In the second stage, we estimate robust covariance matrix minimizing vector variance to label high dimensional outliers. In this stage, we use the robust estimator on the lower-dimensional data space to identify the suspected anomolous observations. The simulation experiments reveal that theintroduced algorithm has a good performance to identify an anomalous observation hidden in a moderate, high, and very high percentage of contamination data and it seems to work well in data analysis.