Title: Selection of unsupervised classification methods for functional data
Authors: Lucas Fernandez Piana - Universidad de Buenos Aires and CONICET (Argentina) [presenting]
Sergi Gonzalez - Agencia Estatal de Meteorologia (Spain)
Ana Justel - Universidad Autonoma de Madrid (Spain)
Julio Rodriguez-Puerta - Universidad Autonoma de Madrid (Spain)
Marcela Svarc - Universidad de San Andres (Argentina)
Abstract: In cluster analysis there is neither an accepted common criterion to evaluate the performance of different procedures, nor a lower bound that indicates the difficulty of the problem. In this context, it is important to develop criteria to compare different clustering procedures on the same data set. The aim is to propose criteria for comparing different partitions in the context of functional data. Inspired by the classification problem, where the methods are easily comparable by the misclassification rate, we build a confusion matrix based on local and global depths. The method is applied to clusters of trajectories and back-trajectories arriving to the Byers Peninsula, located at the western coast of the Livingston Island (South Shetland Islands, Antarctica). Ten years of 5-day back trajectories each six hours, computed with the Hybrid Single-Particle Lagrangian Integrated Trajectory model (HYSPLIT) are analyzed.