GraNDAGAlgorithm<T>

Models & Types · AiDotNet.CausalDiscovery.DeepLearning

GraN-DAG — Gradient-based Neural DAG Learning.

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

For Beginners

GraN-DAG trains a separate neural network for each variable to predict it from the others. The “importance” of each input connection tells us the causal strength, while a mathematical constraint ensures no circular causation.

How It Works

GraN-DAG parameterizes each structural equation f_j as a neural network with sigmoid activations. The weighted adjacency A[i,j] is the PATH PRODUCT of absolute weights across the MLP layers — for the 1-hidden-layer network predicting x_j, A[i,j] = Σ_k |W1_j[i,k]|·|W2_j[k]| (Lachapelle et al. 2020 §2.2), so an input→hidden weight only contributes if that hidden unit actually drives the output. The NOTEARS acyclicity constraint h(A) = tr(e^(A*A)) - d is enforced via augmented Lagrangian.

Reference: Lachapelle et al. (2020), “Gradient-Based Neural DAG Learning”, ICLR.

Constructors

Constructor Summary
GraNDAGAlgorithm(CausalDiscoveryOptions options)

Properties

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