BayesDAGAlgorithm<T>

Models & Types · AiDotNet.CausalDiscovery.Bayesian

BayesDAG — Bayesian DAG learning with gradient-based posterior inference.

Inheritance: CausalDiscoveryBase<T> → BayesianCausalBase<T> → BayesDAGAlgorithm<T>

For Beginners

BayesDAG is a modern Bayesian method that efficiently explores the space of possible causal graphs using gradient-based optimization, providing principled uncertainty quantification about the causal structure.

How It Works

BayesDAG uses a DAG-constrained variational framework that maintains a continuous relaxation of the DAG posterior. It parameterizes edge probabilities via logits Z and uses the Gumbel-Sigmoid trick for differentiable sampling. The acyclicity constraint is enforced via augmented Lagrangian on the expected adjacency.

Algorithm:

  • Initialize edge logits Z[i,j] = 0 (uniform prior on each edge)
  • Sample adjacency: A[i,j] ~ Gumbel-Sigmoid(Z[i,j] / tau)
  • Compute data likelihood: L = -0.5 * ||X - X*A||^2 / n
  • Compute acyclicity: h(A) = tr(e^(A*A)) - d
  • Compute ELBO = likelihood - KL(q || prior) - lambda*h(A)
  • Update Z via gradient ascent on ELBO
  • Anneal temperature tau from soft to hard
  • Threshold final sigmoid(Z) to get binary adjacency

Reference: Annadani et al. (2024), “BayesDAG: Gradient-Based Posterior Inference for Causal Discovery”, NeurIPS.

Constructors

Constructor Summary
BayesDAGAlgorithm(CausalDiscoveryOptions options)

Properties

Property Summary
bool SupportsNonlinear { get; }
string Name { get; }