Måns Karlsson: Trait-based classification of sloppy observations, with regard to error cost

Abstract: The problem at hand will be Bayesian classification of observations based on observed trait- and covariate vectors. We will set up a Bayesian model for classification, which allows for mean and variance parameters to depend on a covariate vector. The model will be expanded to handle a mixture of continous, count, and ordered categorical elements of the trait vector, and we look at how to handle these types of data and missing observations uniformly. Moreover, we construct two particular reward functions, and derive decision rules from these, to handle classification of new observations zealously, or economically. Finally, we highlight the usefulness of this particular model type for taxon identification.