Metrics for evaluating performance and uncertainty of Bayesian network models

This paper presents a selected set of existing and new metrics for gauging Bayesian network model performance and uncertainty. Selected existing and new metrics are discussed for conducting model sensitivity analysis (variance reduction, entropy reduction, case file simulation); evaluating scenarios (influence analysis); depicting model complexity (numbers of model variables, links, node states, conditional probabilities, and node cliques); assessing prediction performance (confusion tables, covariate- and conditional probability-weighted confusion error rates, area under receiver operating characteristic curves, k-fold cross-validation, spherical payoff, Schwarz’ Bayesian information criterion, true skill statistic, Cohen's kappa); and evaluating uncertainty of model posterior probability distributions (Bayesian credible interval, posterior probability certainty index, certainty envelope, Gini coefficient). Examples are presented of applying the metrics to 3 real-world models of wildlife population analysis and management. Using such metrics can vitally bolster model credibility, acceptance, and appropriate application, particularly when informing management decisions.