BCDNetsAlgorithm<T>
Models & Types · AiDotNet.CausalDiscovery.Bayesian
BCD-Nets — Bayesian Causal Discovery Networks.
Inheritance: CausalDiscoveryBase<T> → BayesianCausalBase<T> → BCDNetsAlgorithm<T>
For Beginners
BCD-Nets learn both the graph structure AND the strength of each connection simultaneously, using modern deep learning optimization techniques. They provide uncertainty estimates for both.
How It Works
BCD-Nets use variational inference to approximate the joint posterior over DAG structures and parameters. The graph structure is parameterized via Gumbel-Softmax continuous relaxation of binary edge variables, and parameters (edge weights) are modeled via a factorized Gaussian variational posterior. Both are optimized jointly via gradient ascent on the ELBO.
Algorithm:
- Initialize edge logits Z[i,j] and weight means mu[i,j], log-variances logvar[i,j]
- Sample edges: E[i,j] ~ GumbelSigmoid(Z[i,j]/tau)
- Sample weights: W[i,j] ~ N(mu[i,j], exp(logvar[i,j]))
- Effective adjacency: A = E * W (element-wise)
- Compute ELBO = E_q[log p(X|A)] - KL(q(Z)||p(Z)) - KL(q(W)||p(W)) - lambda*h(E)
- Update Z, mu, logvar via gradient ascent
- Anneal temperature tau and update augmented Lagrangian
Reference: Cundy et al. (2021), “BCD Nets: Scalable Variational Approaches for Bayesian Causal Discovery”, NeurIPS.
Constructors
| Constructor | Summary |
|---|---|
BCDNetsAlgorithm(CausalDiscoveryOptions options) |
Properties
| Property | Summary |
|---|---|
bool SupportsNonlinear { get; } |
|
string Name { get; } |