Rank-based dimension reduction for many-criteria populations

Abstract

Interpreting individuals described by a set of criteria can be a difficult task when the number of criteria is large. Such individuals can be ranked, for instance in terms of their average rank across criteria as well as by each distinct criterion. We therefore investigate criteria selection methods which aim to preserve the average rank of individuals but with fewer criteria. Our experiments show that these methods perform effectively, identifying and removing redundancies within the data, and that they are best incorporated into a multi-objective algorithm.