A clustering technique is described, the objective of which is to detect deviant subpopulations which deviate from a primary subpopulation in1 a well defined direction. The distributions of the primary and deviant subpopulations and the number of deviant subpopulations are initially unknown. Each deviant subpopulation is to be identified by means of a linear scale (an affine function of the observed variables) which distinguishes its members from those of the primary subpopulation. The approach consists of the following three steps: (1) the main cluster of sample points is identified; (2) the directions are determined in which there are the greatest number of outliers from the main cluster; (3) the statistical stability of each supposed outlier direction is tested. Our algorithm has been tested on both manufactured data and on real (MMPI) data.
No takes yet. Share an insight, caveat, or question.
Eric Gold (1976) studied this question.