ICALiNGAMAlgorithm<T>

Models & Types · AiDotNet.CausalDiscovery.Functional

ICA-LiNGAM — Linear Non-Gaussian Acyclic Model using Independent Component Analysis.

Inheritance: CausalDiscoveryBase<T> → FunctionalBase<T> → ICALiNGAMAlgorithm<T>

For Beginners

If you have data generated by a linear causal model with noise that isn’t perfectly bell-shaped (non-Gaussian), ICA-LiNGAM can figure out the complete causal structure — not just which edges exist, but which direction they go.

How It Works

ICA-LiNGAM exploits the identifiability of the linear model X = BX + e when the noise e is non-Gaussian. Using ICA (Independent Component Analysis), it recovers the mixing matrix and derives the causal ordering and edge weights.

Key insight: In a linear Gaussian model, the causal direction is not identifiable from observational data alone. But with non-Gaussian noise, the ICA decomposition uniquely determines the causal structure — this is the LiNGAM identifiability result.

Algorithm:

  • Standardize data to zero mean, unit variance
  • Apply ICA (FastICA) to recover independent components S and mixing matrix A
  • Recover B = I - inv(A) (causal adjacency matrix)
  • Find a permutation that makes B lower-triangular (causal order)
  • Prune small coefficients

Reference: Shimizu et al. (2006), “A Linear Non-Gaussian Acyclic Model for Causal Discovery”, Journal of Machine Learning Research.

Constructors

Constructor Summary
ICALiNGAMAlgorithm(CausalDiscoveryOptions options) Initializes ICA-LiNGAM with optional configuration.

Properties

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