RCDAlgorithm<T>

Models & Types · AiDotNet.CausalDiscovery.Functional

RCD (Repetitive Causal Discovery) — LiNGAM extension for latent confounders.

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

For Beginners

RCD is like DirectLiNGAM but more cautious — it checks whether variables are truly “root causes” or might be influenced by hidden (unobserved) factors. When it finds variables that can’t be cleanly separated, it marks them as potentially confounded rather than forcing a possibly wrong causal direction.

How It Works

RCD extends DirectLiNGAM to handle latent (unobserved) confounders by iteratively identifying “exogenous-like” variables whose residuals are mutually independent, regressing them out, and repeating on the remaining variables.

Algorithm:

  • Standardize the data
  • Initialize the set of remaining variables U = {0, 1, …, d-1}
  • Repeat until U is empty:
  • For each variable i in U, compute residuals after regressing out all discovered ancestors
  • Compute pairwise independence (via mutual information) of residuals
  • Identify the variable whose residuals are most independent of all others (exogenous)
  • Add that variable to the causal ordering and record its causal coefficients
  • Regress out the identified variable from all remaining variables
  • If no sufficiently independent variable found, flag remaining as confounded and stop

Reference: Maeda and Shimizu (2020), “RCD: Repetitive Causal Discovery of Linear Non-Gaussian Acyclic Models with Latent Confounders”, AISTATS.

Constructors

Constructor Summary
RCDAlgorithm(CausalDiscoveryOptions options) Initializes RCD with optional configuration.

Properties

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