Modeling and Forecasting Time Series of Compositional Data: A Generalized Dirichlet Power Steady Model

摘要

This paper presents GDPSM a power steady model (PSM) based on generalized Dirichlet observations for modeling and predicting compositional time series. The model's unobserved states evolve according to the generalized Dirichlet conjugate prior distributions. The observations' distribution is transformed into a set of Beta distributions each of which is re-parametrized as a unidimensional Dirichlet in its exponential form. We demonstrate that dividing the modeling problem into multiple smaller problems leads to more accurate predictions. We evaluate this model with the web service selection application. Specifically, we analyze the proportions of the quality classes that are assigned to the web services interactions. Our model is compared with another PSM that assumes Dirichlet observations. The experiments show promising results in terms of precision errors and standardized residuals.