AmortizedCDAlgorithm<T>

Models & Types · AiDotNet.CausalDiscovery.DeepLearning

Amortized Causal Discovery — meta-learning approach to causal structure learning.

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

For Beginners

Instead of running a slow algorithm each time you have new data, this approach trains a neural network to recognize causal patterns from data statistics. The network learns to map correlation/partial-correlation features to edge probabilities.

How It Works

Amortized Causal Discovery trains a neural network encoder on the input dataset to learn a mapping from data statistics → edge probabilities. The encoder processes pairwise sufficient statistics (covariance, partial correlations) through an MLP to produce edge logits, which are then refined with NOTEARS acyclicity constraint.

Algorithm:

  • Compute sufficient statistics: covariance matrix and partial correlations
  • For each pair (i,j): encode statistics into features via shared MLP
  • Edge logits = MLP output, edge probabilities = sigmoid(logits)
  • Refine with NOTEARS acyclicity penalty via augmented Lagrangian
  • Train encoder end-to-end to minimize reconstruction loss + acyclicity
  • Threshold final edge probabilities

Reference: Lowe et al. (2022), “Amortized Causal Discovery: Learning to Infer Causal Graphs from Time-Series Data”, CLeaR.

Constructors

Constructor Summary
AmortizedCDAlgorithm(CausalDiscoveryOptions options)

Properties

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