AVICIAlgorithm<T>

Models & Types · AiDotNet.CausalDiscovery.DeepLearning

AVICI — Amortized Variational Inference for Causal Discovery.

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

For Beginners

AVICI is like a “universal causal discovery engine” powered by a transformer (similar to ChatGPT’s architecture). It processes data statistics through attention mechanisms to recognize causal patterns.

How It Works

AVICI uses a transformer-inspired architecture to perform causal discovery. The model processes data via self-attention over variable pairs to produce edge probabilities. It computes scaled dot-product attention between variable-pair representations derived from sufficient statistics, then refines edge logits through multiple attention layers.

Algorithm:

  • Compute pairwise features from data: covariance, correlation, variances
  • Project features to query/key/value via learned matrices
  • Apply multi-head self-attention over variable pairs
  • Output layer produces edge logits from attended representations
  • Apply NOTEARS acyclicity constraint via augmented Lagrangian
  • Threshold edge probabilities for final DAG

Reference: Lorch et al. (2023), “Amortized Inference for Causal Structure Learning”, NeurIPS.

Constructors

Constructor Summary
AVICIAlgorithm(CausalDiscoveryOptions options)

Properties

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