DiBSAlgorithm<T>
Models & Types · AiDotNet.CausalDiscovery.Bayesian
DiBS — Differentiable Bayesian Structure Learning.
Inheritance: CausalDiscoveryBase<T> → BayesianCausalBase<T> → DiBSAlgorithm<T>
For Beginners
DiBS uses gradient-based optimization (like training a neural network) to find not just one graph but a whole set of plausible graphs, giving you uncertainty estimates about the causal structure.
How It Works
DiBS uses Stein Variational Gradient Descent (SVGD) to maintain a set of particles, each representing a possible DAG structure via continuous edge logits. The particles are jointly optimized to approximate the posterior distribution over DAGs using a differentiable relaxation of the acyclicity constraint.
Algorithm:
- Initialize K particles, each with edge logits Z_k[i,j] (d x d matrix)
- For each particle, compute edge probabilities via sigmoid(Z_k / tau)
- Compute log-posterior gradient: data likelihood + prior + acyclicity penalty
- Compute SVGD kernel: k(Z_k, Z_l) = exp(-||Z_k - Z_l||^2 / (2*bandwidth^2))
- Update particles: Z_k += lr * (1/K) * sum_l [k(Z_l, Z_k) * grad_l + grad_k(Z_l, Z_k)]
- Anneal temperature tau
- Average edge probabilities across particles for final output
Reference: Lorch et al. (2021), “DiBS: Differentiable Bayesian Structure Learning”, NeurIPS.
Constructors
| Constructor | Summary |
|---|---|
DiBSAlgorithm(CausalDiscoveryOptions options) |
Properties
| Property | Summary |
|---|---|
bool SupportsNonlinear { get; } |
|
string Name { get; } |