The authors propose a new method for Bayesian causal inference, combining order-based MCMC structure learning with gradient-based graph learning. The problem of inferring causal structure is decomposed into inferring a topological order over variables and inferring the parent sets for each variable. The authors use Gaussian processes to model the unknown causal mechanisms, allowing their exact marginalisation. Rao-Blackwellisation is introduced, where all components except the causal order are eliminated. The proposed method shows superior results on linear and non-linear additive noise benchmarks with scale-free and Erdos-Renyi graph structures.
Publication date: 22 Feb 2024
Project Page: https://arxiv.org/abs/2402.14781v1
Paper: https://arxiv.org/pdf/2402.14781
