Title: Asymmetric projection pursuit for classification
Authors: Nicola Loperfido - University of Urbino (Italy) [presenting]
Abstract: When applied to cluster analysis, projection pursuit aims at finding the data projections which best separate clusters. Asymmetric projection pursuit for classification finds the direction which best separates one cluster from the remaining data, removes the cluster and repeats the previous steps until no clusters are left. The chosen projection pursuit index is kurtosis, which is particularly apt at data dichotomization. Asymmetric projection pursuit builds upon asymmetric linear dimension reduction for classification. The method is theoretically motivated using finite normal mixtures and it is empirically illustrated with the Iris dataset.