Effective Bayesian Causal Inference via Structural Marginalisation and Autoregressive Orders

Abstract

Bayesian causal inference (BCI) naturally incorporates epistemic uncertainty about the true causal model into down-stream causal reasoning tasks by posterior averaging over causal models. However, this poses a tremendously hard computational problem due to the intractable number of causal structures to marginalise over. In this work, we decompose the structure learning problem into inferring (i) a causal order and (ii) a parent set for each variable given a causal order. By limiting the number of parents per variable, we can exactly marginalise over the parent sets in polynomial time, which leaves only the causal order to be marginalised. To this end, we propose a novel autoregressive model over causal orders (ARCO) learnable with gradient-based methods. Our method yields state-of-the-art in structure learning on simulated non-linear additive noise benchmarks with scale-free and Erdos-Renyi graph structures, and competitive results on real-world data. Moreover, we illustrate that our method accurately infers interventional distributions, which allows us to estimate posterior average causal effects and many other causal quantities of interest.