Abstract

This paper proposes a model learning Semi-parametric rela- tionships in anExpert Bayesian Network (SEBN) with linear parameter and structure constraints.We use Gaussian Pro- cesses and a Horseshoe prior to introduce minimal nonlin-ear components. To prioritize modifying the expert graph over adding new edges,we optimize differential Horseshoe scales. In real-world datasets with unknowntruth, we gen- erate diverse graphs to accommodate user input, addressingidentifiability issues and enhancing interpretability. Evalua- tion onsynthetic and UCI Liver Disorders datasets, using metrics like structuralHamming Distance and test likelihood, demonstrates our models outperformstate-of-the-art semi- parametric Bayesian Network model.

 

Quick Read (beta)