IGCIAlgorithm<T>
Models & Types · AiDotNet.CausalDiscovery.Functional
IGCI (Information-Geometric Causal Inference) — bivariate causal discovery via entropy.
Inheritance: CausalDiscoveryBase<T> → FunctionalBase<T> → IGCIAlgorithm<T>
For Beginners
IGCI works by checking whether the relationship between two variables is “smoother” in one direction than the other. If X causes Y, the mapping from X to Y tends to have slopes that anti-correlate with the density of X — a natural geometric property of cause-effect relationships. This method is very fast (pairwise comparisons only) but only works for bivariate, deterministic-ish relationships.
How It Works
IGCI determines causal direction between pairs of variables by exploiting the information-geometric principle: if X causes Y, the marginal distribution of X and the conditional P(Y|X) are independent, which means the slope of Y=f(X) tends to correlate negatively with the density of X at that point.
Algorithm:
- For each pair (X, Y), uniformize both variables to [0,1] via rank transform
- Sort by X values and compute the average log-slope: sum(log|dy/dx|)/n
- The causal direction score is: score = mean(log|f’(x)|)
- If score < 0, X→Y is preferred; if score > 0, Y→X is preferred
- Build adjacency matrix from pairwise decisions
Reference: Janzing et al. (2012), “Information-Geometric Approach to Inferring Causal Directions”, Artificial Intelligence.
Constructors
| Constructor | Summary |
|---|---|
IGCIAlgorithm(CausalDiscoveryOptions options) |
Initializes IGCI with optional configuration. |
Properties
| Property | Summary |
|---|---|
bool SupportsNonlinear { get; } |
|
string Name { get; } |