A1292
Title: Bayesian presence-only data modeling validation
Authors: Giovanna Jona Lasinio - Sapienza University of Rome (Italy) [presenting]
Gian Mario Sangiovanni - Sapienza University (Italy)
Gianluca Mastrantonio - Politecnico of Turin (Italy)
Abstract: Motivated by the need for a reliable assessment of predictive performance in spatial point process models, a revisiting of cross-validation strategies for spatially dependent observations is presented. Standard random-assignment or multinomial thinning schemes are computationally convenient, but they ignore spatial autocorrelation and often produce overly optimistic error estimates because the training and validation sets remain strongly dependent. The presentation, therefore, argues for spatial block cross-validation, whose goal is to evaluate genuinely out-of-sample, longer-range prediction rather than short-range interpolation. It draws on leave-group-out cross-validation ideas developed for latent Gaussian models, where groups are constructed to remove the observations most informative for predicting a held-out point, using correlations in the linear predictor. Building on this motivation, the proposed direction for point processes is to define spatially meaningful blocks using a graph constructed from a reference set, and then apply clustering to obtain folds that better reflect the underlying dependence structure. This creates a principled compromise between preserving spatial structure and maintaining enough separation for honest validation. The talk also highlights open methodological questions.
