B1093
Title: The use of Riordan arrays for the hyperparameter choice of prior distributions with consistent EPPFs
Authors: Jan Greve - WU Vienna University of Economics and Business (Austria) [presenting]
Abstract: In Bayesian clustering based on mixture models, prior distributions with the exchangeable partition probability function (EPPF) equipped with consistency are often utilized. In particular, Gibbs-type priors with a multiplicative EPPF have seen uses in many applications. These priors are by construction biased such that any hyperparameter choice will result in the concentration of the majority of the probability masses to a small subset of the entire support of the distribution. The use of the Riordan array is proposed, a recent tool in combinatorics, to characterize the biasedness of such prior distributions to aid appropriate hyperparameter choice. In addition, the computational efficiency of the approach is compared to the algorithm based on recursion.