GFCIAlgorithm<T>

Models & Types · AiDotNet.CausalDiscovery.Hybrid

GFCI — Greedy FCI, a hybrid of GES and FCI.

Inheritance: CausalDiscoveryBase<T> → HybridBase<T> → GFCIAlgorithm<T>

For Beginners

GFCI is useful when you suspect there are hidden variables affecting your data. It first quickly finds a good graph structure, then checks whether some connections might actually be due to hidden common causes rather than direct effects.

How It Works

GFCI combines score-based and constraint-based approaches to handle latent confounders. Phase 1 uses greedy score-based search (like GES) to find an initial skeleton and orientations. Phase 2 applies FCI-like rules to detect possible latent confounders and convert directed edges to bidirected edges where appropriate.

Algorithm:

  • Score-based phase: Use greedy hill climbing with BIC to find an initial DAG
  • Skeleton validation: Test remaining edges with CI tests; remove if independent
  • Latent detection: For each unshielded triple i — k — j where i and j are

non-adjacent, check if k is in the separation set. If not, orient as collider (i → k ← j)

  • Discriminating path check: Identify possible bidirected edges (latent confounders)

by checking if edges oriented in both directions can be explained by a common cause

  • Apply FCI orientation rules to propagate orientations

Reference: Ogarrio et al. (2016), “A Hybrid Causal Search Algorithm for Latent Variable Models”, PGM.

Constructors

Constructor Summary
GFCIAlgorithm(CausalDiscoveryOptions options) Initializes GFCI with optional configuration.

Properties

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