DECIAlgorithm<T>

Models & Types · AiDotNet.CausalDiscovery.DeepLearning

DECI — Deep End-to-end Causal Inference.

Inheritance: CausalDiscoveryBase<T> → DeepCausalBase<T> → DECIAlgorithm<T>

For Beginners

DECI simultaneously learns “which variables cause which” and “how they cause each other.” It’s particularly good at handling complex, non-standard relationships and can also estimate intervention effects.

How It Works

DECI is a flow-based variational inference method that jointly learns the causal graph and the functional relationships between variables. It uses normalizing flows to model flexible conditional distributions and a variational distribution over DAGs.

Algorithm:

  • Initialize edge logits Z[i,j] and per-variable MLPs (location networks)
  • Sample soft adjacency: A[i,j] = sigmoid(Z[i,j] / tau) with temperature annealing
  • For each target j: predict location f_j = MLP_j(masked input by A[:,j])
  • Compute log-likelihood using Gaussian noise model: log p(x_j | pa(j))
  • Add KL divergence for edge distribution and NOTEARS acyclicity penalty
  • Optimize ELBO via gradient descent on Z and MLP weights
  • Threshold final edge probabilities to get DAG

Reference: Geffner et al. (2022), “Deep End-to-end Causal Inference”, arXiv.

Constructors

Constructor Summary
DECIAlgorithm(CausalDiscoveryOptions options)

Properties

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