Bayesian Categorical Modeling of Multidimensional QoE in Haptic-Audiovisual Communications

Abstract

This paper proposes a methodology for Bayesian modeling of multidimensional ordinal categorical data for QoE by means of the structural equation modeling (SEM) technique, focusing on the identifiability (indeterminacies) problem of latent variables. As an example material, we adopt a haptic-audiovisual interactive communication system along with previously collected data of 13 subjective QoE measures and a single objective QoE measure. To reduce the dimensionality of the model, we introduce two exogenous latent constructs, Audiovisual quality (AVQ) and Haptic quality (HQ), and an endogenous construct User experience quality (UXQ). For modeling of the categorical data of scores, we also employ continuous latent variables underlying the scores. We resort to the unit loading (UL) method and the unit variance (UV) method for the resolution of scale indeterminacies. Bayesian estimates of free parameters including the loading and regression coefficients by the two methods have different values, while estimates of the QoE measures have almost the same values, which are found to be satisfactory fit in terms of the overall posterior predictive p-values. A comparison of the two methods in terms of the p-value and DIC demonstrates that the model with UL constraints is somewhat better than the model with UV. We also learn that HQ has much stronger effects on UXQ than AVQ regardless of the identifiability constraint (UL or UV).