CCDrAlgorithm<T>
Models & Types · AiDotNet.CausalDiscovery.Functional
CCDr (Concave penalized Coordinate Descent with reparameterization) for DAG learning.
Inheritance: CausalDiscoveryBase<T> → FunctionalBase<T> → CCDrAlgorithm<T>
For Beginners
CCDr finds causal relationships by fitting a statistical model where each variable depends on a sparse set of “parent” variables. It uses a clever penalty (MCP) that encourages most connections to be zero while keeping strong ones unbiased. The algorithm checks each potential connection one at a time and only keeps it if removing it would significantly hurt the model fit.
How It Works
CCDr learns a Bayesian network structure by minimizing a penalized negative log-likelihood using coordinate descent with a concave penalty (MCP — Minimax Concave Penalty). The MCP penalty provides near-unbiased estimation of large coefficients while still inducing sparsity.
Algorithm:
- Initialize adjacency matrix W = 0
- For each target variable j, cycle through candidate parents i != j
- Update W[i,j] via coordinate descent on the penalized Gaussian log-likelihood
- Apply soft-thresholding with MCP penalty derivative
- Enforce acyclicity: reject updates that would create a cycle
- Repeat until convergence or max iterations
Reference: Aragam and Zhou (2015), “Concave Penalized Estimation of Sparse Gaussian Bayesian Networks”, JMLR.
Constructors
| Constructor | Summary |
|---|---|
CCDrAlgorithm(CausalDiscoveryOptions options) |
Initializes CCDr with optional configuration. |
Properties
| Property | Summary |
|---|---|
bool SupportsNonlinear { get; } |
|
string Name { get; } |