Title: Dynamic mixture of experts models for online predictions
Authors: Parfait Munezero - Stockholm University and Ericsson AB (Sweden) [presenting]
Mattias Villani - Stockholm University (Sweden)
Abstract: Mixture of experts models provide a flexible framework of modelling the density of a response variable using finite mixture models with component density functions and mixture weights depending on a set of covariates. We propose a class of dynamic mixture of experts models for online (real-time) predictions. Our model allows the component models to be any density functions not necessarily limited to the exponential family, and the parameters to vary over time. The inference is done in a Bayesian framework using sequential Monte Carlo (SMC) a.k.a Particle filter algorithms. The smoothness of parameters is controlled through a random walk prior process, which allows the parameters to fluctuate and adapt locally through time or to remain constant over time. We, therefore, propose an inference method that applies to models with either static or dynamic parameters in a unified way. We apply the model to a real dataset consisting of faults reported on a series of upgrades of a large-scale software. Further, we assess the performance of our inference method using different simulation scenarios.