Reproduce “Bayesian learning of Causal Structure and Mechanisms with GFlowNets and Variational Bayes” on OpenResearch

About this paper

Mizu Nishikawa-Toomey, Tristan Deleu, Jithendaraa Subramanian, Yoshua Bengio, Laurent Charlin

Variational Bayes-DAG-GFlowNet (VBG) introduces a method to jointly learn the posterior distribution over causal graph structures and their mechanisms, overcoming limitations of previous graph-only or acyclicity-constrained methods. The approach demonstrates competitive performance in recovering graph structures and mechanism parameters on synthetic and real-world datasets, particularly for smaller to medium-sized graphs.

Key insights

• Combining GFlowNets with Variational Bayes provides a robust framework for jointly inferring causal graphs and their mechanisms, ensuring acyclicity by design throughout training and sampling.
• The use of a factorized variational distribution and iterative ELBO maximization enables efficient learning of both structural and parametric uncertainty in a complex, high-dimensional space.
• The framework is conceptually flexible enough to accommodate different types of causal mechanisms beyond linear Gaussian models, provided the 'delta score' (change in log-reward for an edge addition) can be efficiently computed.

Summary by alphaXiv.