Accurate logistic variational message passing: algebraic and numerical details

We provide full algebraic and numerical details required for fitting accurate logistic likelihood regression-type models via variational message passing with factor graph fragments. Existing methodology of this type involves the Jaakkola-Jordan device, which is prone to poor accuracy. We examine two alternatives: the Saul-Jordan tilted bound device and conjugacy enforcement via multivariate normal prespecification of a key message. Both of these approaches appear in related literature. Our contributions facilitate immediate implementation within variational message passing schemes. Copyright (C) 2017 John Wiley & Sons, Ltd.